Anthropic raised $65 billion and is now worth more than OpenAI.

Project Glasswing reported its first month of results, and the numbers are extraordinary and alarming in equal measure.

The Pope published the first papal encyclical on artificial intelligence.

And the White House quietly approved $9 billion to stop US spy agencies from falling behind on AI chips.

Here’s everything you need to know before Monday gets the best of you.

Anthropic Hit $965 Billion. The October IPO Is Now A Question Of When, Not If.

On May 28, Anthropic closed a $65 billion Series H at a post-money valuation of $965 billion, passing OpenAI’s $852 billion mark for the first time.

The round was led by Altimeter, Dragoneer, Greenoaks, and Sequoia, with participation from D.E. Shaw, Blackstone, and DST Global.

The figure includes $15 billion of previously committed hyperscaler investment, including $5 billion from Amazon. The round moved from first investor conversations in late April to close in under five weeks.

The revenue picture. Anthropic reported a $47 billion revenue run rate on Thursday, up from $9 billion annualised at the end of 2025 and $30 billion in April.

Gross margins moved from 38 percent a year ago to over 70 percent. Eight of the Fortune 10 are active clients. Over 1,000 enterprise accounts exceed $1 million in annual spend.

The raise is structured to prepay for compute ahead of a planned IPO.

The IPO timeline. Multiple banks are in preliminary discussions. Bloomberg reports a listing could arrive as early as October 2026.

The valuation trajectory shows how fast that window is closing: $380 billion in February, $900 billion in May, $965 billion post-round on May 28.

An October listing at these levels would be the largest technology IPO since at least 2014.

Why it matters

Anthropic tripled its valuation in three months on revenue that tripled in the same period. That is not valuation inflation. That is a business compounding faster than almost anything in software history.

Project Glasswing Found 23,000 Vulnerabilities. Less Than 1% Are Patched.

Anthropic published its first monthly report for Project Glasswing on May 22.

The initiative, powered by the unreleased Claude Mythos Preview model, has helped partners identify over 10,000 high- and critical-severity software vulnerabilities in one month.

The fuller picture: Mythos scanned over 1,000 open-source projects and flagged 23,019 vulnerabilities. Of those, 6,202 were estimated to have high or critical severity.

Independent security firms checked a 1,752-finding sample and confirmed 90.6% were real bugs.

Cloudflare found 2,000 bugs, 400 of them high or critical. Mozilla found and fixed 271 vulnerabilities in Firefox, 10 times more than it found using an older Claude model.

The capability picture. Palo Alto Networks said Mythos accomplished the equivalent of a year of pentesting in under three weeks, and noted impressive vulnerability-chaining — combining medium- and low-severity issues into critical exploits.

One wolfSSL flaw, CVE-2026-5194, would have let attackers forge TLS certificates across billions of IoT and industrial devices. A banking partner stopped a $1.5 million fraudulent wire transfer mid-execution.

Microsoft said patch releases will “continue trending larger for some time,” citing Mythos discoveries in part.

The number that should worry security teams. Less than 1% of the vulnerabilities Mythos found have been patched.

A model that finds bugs faster than human teams can respond has produced an enormous backlog of known vulnerabilities that bad actors could also find if they run a comparable model.

Anthropic plans to expand Glasswing with US and allied government partners, and intends to make Mythos-class models generally available once stronger safeguards exist. The patching problem needs to be solved first.

Why it matters

Finding 23,000 vulnerabilities is extraordinary. Patching less than 1% is dangerous. AI can discover bugs at a scale humans cannot match. The rest of the security process, triage, remediation, and disclosure, was not built for this pace.

The Pope Wrote The First Encyclical On AI. An Anthropic Co-Founder Spoke At The Vatican.

On May 25, Pope Leo XIV released Magnifica Humanitas, the first encyclical of his papacy and the first major papal document ever written specifically about artificial intelligence.

The title translates as “Magnificent Humanity.” The Pope signed it on May 15, the 135th anniversary of Rerum Novarum, the encyclical Leo XIII wrote about industrial-era workers in 1891.

The underlying premise: technology is not “a force antagonistic to humanity” and not “inherently evil,” but “technology is never neutral, because it takes on the characteristics of those who devise, finance, regulate, and use it.”

The presentation. Leo XIV presented the encyclical personally at the Vatican, unlike most popes who delegate the task to cardinals.

AI experts attended, including Anthropic co-founder Chris Olah. Olah, who is not a believer, issued a call to religious communities, civil society, academics, and governments to follow the Pope’s example: “to take this seriously, to look closely, and to push events in a better direction. We need informed critics who will tell the labs when we are failing.”

The argument. Magnifica Humanitas pushes back on a specific strand of AI optimism: the idea that human limitations, illness, ageing, and suffering are defects to be optimised away.

It argues humans often flourish through their limitations, and that AI should support what the document calls “openness and communion.”

The text does not call for a ban or moratorium. It calls for governance frameworks that ensure AI serves the common good and does not concentrate power in ways that exclude the most vulnerable.

The comparison to Leo XIII and industrial-era workers is deliberate: the Church arrived late to that transformation. The document is a statement that it intends to arrive earlier this time.

Why it matters

The Catholic Church has 1.3 billion members and a history of social teaching that has shaped labour law, human rights frameworks, and political philosophy across two centuries. Magnifica Humanitas is the institution formally stating that AI governance is a moral question, not just a technical one. That carries weight in rooms where most AI companies do not have a seat.

The White House Quietly Approved $9 Billion To Buy US Spy Agencies Out Of A Chip Shortage.

The New York Times reported on May 23 that the White House has approved a secret $9 billion request to acquire frontier AI chips for US intelligence agencies.

New AI models use more compute than even most technology experts anticipated a year or two ago. That fuelled concerns in the White House and Congress that a chip shortage was causing intelligence agencies to fall behind on top-secret AI work.

The $9 billion is intended in part to boost the infrastructure that can support Nvidia’s Grace Blackwell superchip. Congress still must approve the funding, but the administration is reprogramming $800 million for faster acquisition of compute capacity.

The infrastructure picture. The agencies primarily run classified AI models on Amazon Web Services. Amazon announced a $50 billion effort last year to upgrade its government cloud services.

The capability gap is not primarily a model gap but an infrastructure gap. The agencies have access to commercial frontier models. What they lack is dedicated compute to run them at top-secret scale and access controls.

The $9 billion is buying dedicated capacity, not new model capability.

The cancelled order. On Thursday, the White House abruptly cancelled a signing ceremony for a new executive order on artificial intelligence hours before it was scheduled to begin.

President Trump told reporters he “didn’t like aspects of it.” No revised timeline has been announced.

The order had been in development for months and was expected to address AI safety standards, export controls, and AI use in federal agencies.

A secret $9 billion chip approval and a cancelled AI governance order in the same week is the clearest statement of administration priorities: capability first, governance later.

Why it matters

US intelligence agencies running on a chip shortage is a national security gap; the White House is now treating it with the urgency of a conventional weapons shortfall. The cancelled executive order is the other half of that story: the capability investment is moving, and the governance framework is not.

DeepMind Cracked Nine Open Maths Problems. Two Had Been Unsolved For 56 Years.

On May 21, Google DeepMind published an arXiv preprint documenting results from AlphaProof Nexus, a system that pairs a language model with Lean, a formal proof assistant that checks every logical step against mathematical axioms.

AlphaProof Nexus autonomously cracked 9 out of 353 open Erdős problems and proved 44 out of 492 open conjectures from the Online Encyclopedia of Integer Sequences.

Two of the nine Erdős problems had been open for 56 years. The cost was a few hundred dollars per problem.

How it works. The AI proposes a proof. Lean checks every logical step against mathematical axioms. If the proof doesn’t hold, it gets rejected.

This changes the nature of the result. Language models are fluent and often wrong. Lean proofs are checked by a compiler.

One academic mathematician who reviewed both this and a concurrent OpenAI result said, “OpenAI’s result was impressive, but it was a natural language argument that still requires peer review. AlphaProof Nexus outputs Lean proofs. There is no ambiguity about correctness.”

DeepMind CEO Demis Hassabis tempered expectations, saying the system is “still not AGI.”

Why are open problems different? They may have resisted specialists for decades. They do not come with the comfort of knowing a solution exists.

Carnegie Mellon mathematician Jeremy Avigad framed it in a March essay: “We are running out of places to hide. We have to face up to the fact that AI will soon be able to prove theorems better than we can.”

AlphaProof Nexus solved 2.5% of the open Erdős catalogue. The interesting question is whether that becomes 10% next year and 30% the year after.

Why it matters

AI solving open research mathematics with machine-verifiable proofs at a few hundred dollars each is a different category of result from anything before it. Mathematics has been one of the last domains where human originality seemed irreplaceable. That assumption is now under pressure.

NVIDIA Crossed $4 Trillion. Every Story This Week Was Good News For Its Stock.

NVIDIA’s market capitalisation crossed $4 trillion this week, a milestone no company has reached before.

The move was driven by converging demand signals: Anthropic’s $65 billion raise, the White House’s $9 billion AI chip approval, and Microsoft’s pre-Build announcements expanding its AI infrastructure stack.

Each story contains Nvidia hardware as the substrate. Anthropic trains on Nvidia GPUs. The Grace Blackwell superchip is the specific hardware that the $9 billion intelligence budget targets. Microsoft’s AI infrastructure relies heavily on Nvidia.

Almost every major AI development in either the commercial or national security sector is a direct driver of Nvidia’s revenue.

The supply constraint. Blackwell has been in shortage since launch. Every major hyperscaler has disclosed that GPU availability is a binding constraint on its AI revenue.

That constraint is not expected to ease until the next generation ships at scale. NVIDIA is selling into a market where demand structurally exceeds supply for at least the next two quarters.

That is the simplest explanation for a $4 trillion valuation: the world needs more than NVIDIA can produce, and the demand is coming from customers committed to spending hundreds of billions to buy it.

The longer-term risk. Google, Amazon, Meta, Microsoft, and OpenAI are all developing inference-optimised chips designed to run their models without paying NVIDIA’s margins.

None have displaced NVIDIA in training, and none are likely to in the near term. But inference is where production compute volume runs, and the share flowing to custom silicon is growing with each chip generation.

Why it matters

NVIDIA at $4 trillion means the market believes AI infrastructure demand will outpace supply for years. The custom silicon risk is real, but it is a long-cycle story. Right now, the shortage is the story.

Microsoft Build Opens This Week. Windows Just Became An Agent Platform.

Microsoft Build 2026 opens June 2 in San Francisco. The announcements arriving in preview coverage this week signal the largest repositioning of Windows since the cloud era began.

The centrepiece is the Windows Agent Runtime, a background service that manages agent lifecycles, memory, and permissions at the operating system level.

Agents built on it interact directly with Windows subsystems, file manager, task scheduler, and hardware sensors, using declarative APIs rather than screen-scraping or UI automation.

This repositions Windows to compete with cloud-native agent platforms like LangChain, AutoGen, and Semantic Kernel, but with a billion-device install base behind it.

The platform pieces. All WinRT APIs available to the built-in Copilot will also be available to third-party agents that pass the same certification process.

The Azure Agent Mesh federates agent execution across on-premises Windows servers, Windows 365 Cloud PCs, and Azure Arc edge devices, with consumption-based pricing launching in Q4 2026.

Computer-using agents are now available in Copilot Studio. The multi-model Copilot platform, rebuilt to include Anthropic’s Claude alongside OpenAI’s GPT family, is the most visible sign of Microsoft’s shift away from exclusive model dependency.

The shift. Microsoft has moved from “AI features in Windows” to “Windows as an agent execution environment.”

The billion-device install base means whatever agent runtime ships at Build will reach a larger population of devices than any competing platform, immediately.

Why it matters

No other agent platform has Windows’ distribution. Whether developers build on it depends on the APIs being good enough, but the reach is unmatched from day one.

China Just Banned Foreign AI Chips From Government Systems. The Domestic Replacement Plan Is On.

China added AI chips to its state-backed security assessment list this week, formalising a directive that government systems and critical infrastructure must evaluate and ultimately replace foreign AI hardware with domestic alternatives.

The move was framed as a security measure. The practical target is Nvidia hardware, which remains the dominant GPU in Chinese data centres despite successive rounds of US export controls.

The domestic ecosystem. Huawei’s Ascend 910C and 910D have been deployed at scale. DeepSeek’s V4 model, covered two issues ago, was trained and deployed entirely on domestic silicon.

The performance gap between Huawei’s best chips and Nvidia’s current generation remains significant on raw benchmarks. But for government workloads, which tend to be inference-heavy rather than training-heavy, the gap is narrower and running entirely on domestic hardware is feasible.

The signal. Adding AI chips to a security assessment list is not an immediate ban. It is the beginning of a structured replacement process that typically runs three to five years in Chinese government procurement.

Huawei, Cambricon, and Biren Technology now have a mandated market that will grow over that timeline, regardless of how their products perform in open commercial competition.

US export controls designed to slow Chinese AI development have produced a policy response that will accelerate Chinese domestic chip investment through guaranteed government procurement. The intended effect and the actual effect are moving in different directions.

Why it matters

China just created a mandated government market for domestic AI chips. The chips are currently behind. The market is now guaranteed to exist until they catch up.

Claude Opus 4.8 Just Shipped. Anthropic Is Now Running Two Capability Tiers At Once.

Alongside the funding announcement on May 28, Anthropic released Claude Opus 4.8, updating its flagship model with improvements in extended thinking, multi-step agentic tasks, and coding.

The model incorporates safety lessons from Project Glasswing and Claude Mythos Preview, threading defensive cybersecurity research from the restricted model into the commercial one.

Pricing holds at $5 per million input tokens and $25 per million output. The model is available through Claude.ai, the API, Amazon Bedrock, and Vertex AI.

The benchmark picture. Opus 4.8 leads GPT-5.5 on complex multi-step reasoning, SWE-bench Pro issue resolution, and long-context document analysis.

GPT-5.5 retains its lead on Terminal-Bench 2.0 and agentic computer use tasks. Enterprise teams are increasingly routing tasks by model rather than committing to a single vendor.

The cadence. Opus 4.5, 4.6, and 4.7 shipped within three months of each other. Opus 4.8 arrives roughly six weeks after 4.7.

With Mythos-class capability in restricted access and Opus 4.8 now the public frontier, Anthropic is running two distinct capability tiers at commercial scale, something no other lab is currently doing.

Why it matters

The gap between Opus 4.8 and what Mythos can do is the most consequential unknown in the current AI landscape.

And that wraps up this week. Tune in next Monday, same time, for another deep-dive into the stories shaping the AI world.

The Sentinel lands in your inbox every Monday so you can catch up with the fast-moving AI space while sipping your morning coffee. Every detail that matters, none that doesn’t.

Google I/O delivered the biggest AI conference event in years. Three humanoid robots sorted 88,000 packages without stopping while 3 million people watched live. Microsoft’s infrastructure spend crossed $190B. And OpenAI’s custom chip entered production at Samsung.

Here’s everything you need to know before Monday gets the best of you.

Google I/O Wasn’t A Developer Conference. It Was A Declaration Of War.

Over two days in Mountain View on May 19 and 20, Google made more AI announcements than at any event any company has held in the past two years.

CEO Sundar Pichai’s message was direct: AI should not be something you visit. It should be the infrastructure you live inside.

Google backed that with the biggest Search redesign in 27 years, two new flagship Gemini models, a 24/7 personal AI agent, a $100/month AI Ultra tier, Samsung XR smart glasses, and a live demo of an AI building an entire operating system from scratch on stage.

The models. Gemini 3.5 Flash is the new flagship.

Google claims it is four times faster than rival frontier models while matching or exceeding their quality across coding, reasoning, and multimodal tasks. Early developer testing has supported the speed claims.

Gemini Omni Flash is a world model trained on continuous streams of video, audio, and text at once, giving it persistent contextual awareness that single-modality models can’t match.

The pricing move. This is what changes competitive dynamics most directly.

AI Ultra at $100/month gives subscribers access to every Gemini model, Gemini Spark, 20TB of cloud storage, and Gemini in Workspace, with five times higher usage limits than the Pro plan.

ChatGPT Pro costs $200. Claude Max starts at $100. Google has cut the comparable price in half at launch.

That is a market share bet, not a margin bet. Google has the runway to sustain a loss long enough to reshape what AI subscription pricing looks like.

Why it matters

Google launched more consequential AI products in two days than most companies launch in a year, and undercut OpenAI and Anthropic on price at the same time. Willingness to lose money for market share is a different competitive threat than a better model.

Gemini Spark Runs While You Sleep.

Gemini Spark is Google’s “24/7 personal agent for work, school, and daily life,” and it’s the I/O product that deserves its own story.

Every other AI assistant operates on a request-response model. You open the app, send a message, get an answer, and close the app.

Spark breaks that completely. It runs continuously on Google’s cloud, so it keeps working after your phone screen goes dark and your laptop lid is shut.

It has access to your apps, calendar, email, and documents, and takes actions on your behalf without being asked.

A “Daily Brief” feature scans your calendar and inbox overnight and delivers a prioritised plan every morning before you’ve opened anything yourself.

The architecture. Running an agent continuously in the cloud is a different engineering problem from running inference on demand.

Google has built session persistence, context retention across days and weeks, and Workspace integration at a depth that lets Spark act rather than just advise.

The unknowns. Spark is rolling out to AI Ultra subscribers over the next few weeks.

The real questions can only be answered in production: how it handles conflicting priorities, what it does when an autonomous action turns out wrong, and how much retained context proves useful versus noise.

Why it matters

Every AI assistant before Spark waited for you. Spark doesn’t. A persistent cloud agent that operates continuously and takes actions without prompting is a different product category. Whether the reality matches the demo will decide whether this week marks a real shift or an impressive announcement.

Three Humanoid Robots Ran For 40 Hours Straight. 3 Million People Watched.

Figure AI planned an eight-hour demo. Bob, Frank, and Gary kept going for nearly 40.

The company’s Helix-02-powered humanoid robots began a livestreamed package-sorting run at their San Jose warehouse on May 13.

By the time the run passed 24 hours, 3 million people were watching on X, naming the robots and posting updates like a reality show. A fourth robot, Rose, joined mid-run.

Final numbers: four robots sorted 88,000 packages in 72 hours with zero reported failures. The pace was roughly one package every three seconds, matching human warehouse worker averages.

The technical claims. They hold up with caveats.

The robots detect barcodes using onboard cameras, pick packages, and place them barcode-face-down onto a conveyor belt. Everything runs through Helix-02 onboard, with no teleoperation.

Observers noted visible errors: packages placed barcode-up, one knocked off a belt.

CEO Brett Adcock said it directly: “F.03 is now around human parity.” Not better. Parity.

The honest data point. A human vs. robot test during the stream gave the most useful number of the week.

Over an eight-hour shift, the F.03 robot sorted 12,732 packages at 2.83 seconds per item. The human, a Figure employee named Aime, sorted 12,924 packages at 2.79 seconds.

The human won by 192 packages.

The real milestone isn’t speed. It’s that robots can sustain the pace for 40 hours without fatigue, breaks, or shift changes.

For 24-hour operations with rotating human shifts, that endurance changes the economic calculation even without a speed advantage.

Why it matters

A humanoid robot matching human speed at a repetitive logistics task, then running it for 40 hours without stopping, is a production viability data point. Warehousing and logistics are watching closely.

Anthropic’s $30 Billion Round Just Closed. It’s Now The Most Valuable Private Company In History.

Anthropic’s $30 billion funding round closed this week at a pre-money valuation exceeding $900 billion, passing OpenAI’s $852 billion mark from March.

The round was led by Greenoaks, with Sequoia, Altimeter, and Dragoneer each investing $2 billion or more.

CFO Krishna Rao began gauging demand in late April, and the round assembled and closed in under five weeks.

Capital will mainly prepay for compute ahead of a planned IPO, which Bloomberg reports could arrive as early as October 2026.

The revenue trajectory. At the end of 2025, Anthropic’s annualised revenue was approximately $9 billion. By April 2026, it had passed $30 billion. The company projects $45 billion shortly.

Claude Code alone generates $2.5 billion annualised. Eight of the Fortune 10 are active clients. Over 1,000 enterprise accounts exceed $1 million in annual spend.

Gross margin moved from 38 percent a year ago to over 70 percent now.

That answers the standard objection to AI valuations: at 70 percent gross margin, Anthropic is a software business, not an infrastructure business.

Why it matters

Anthropic went from $380 billion to $900 billion in three months on revenue that grew 80x in 18 months and margins that crossed 70 percent. An October IPO at these levels would be among the most significant tech listings since the dot-com era. The company, founded as a safety-focused alternative to OpenAI, is now worth more than OpenAI.

Google Showed Off Its AI Glasses. Samsung Is Building Them. They Ship In The Autumn.

The final segment of Google’s I/O keynote was reserved for hardware.

Google and Samsung teased Android XR smart glasses with two variants: an audio-only model that functions as a continuous Gemini speaker, and a display-lens model that overlays information in the wearer’s field of view.

Both are built around continuous Gemini conversations rather than discrete app interactions.

The Samsung XR glasses ship in autumn 2026. Price unconfirmed.

The context. Google Glass launched in 2013 and failed because it looked strange, felt expensive, and had no compelling daily use case.

The Android XR glasses arrive in a different environment. Modern AI can make ambient, always-on assistance genuinely useful in ways 2013 technology could not.

Cameras and mics in the frame feed real-time context to Gemini. Live translation, contextual navigation, and hands-free information retrieval are obvious early use cases.

The design problem that killed Glass isn’t fully solved. Neither variant is yet consumer-ready in appearance.

But Samsung building the hardware and Google providing the AI is the partnership structure that gives this its best chance.

Why it matters

Wearable AI glasses have failed every time because the AI wasn’t good enough to justify wearing a computer on your face. The AI is now good enough. The question this autumn is whether the hardware is comfortable enough and the use cases obvious enough for consumers.

OpenAI’s Custom Chip Just Entered Production. Samsung Is Making It. NVIDIA Is Watching.

OpenAI’s custom AI chip program, codenamed Project Titan, entered production at Samsung this week.

The chip is co-developed with Broadcom, manufactured on TSMC’s 3nm process, and uses Samsung’s HBM4 memory under an exclusive supply agreement.

The first-generation chip is an inference accelerator, designed to run AI models at scale rather than train new ones.

The milestone matters because OpenAI has relied on NVIDIA for almost all its compute since its founding. Every ChatGPT inference call, every API response, runs on NVIDIA hardware. Project Titan begins changing that.

The economics. OpenAI’s inference costs for reasoning models are one of the largest constraints on its business model.

Reasoning models use substantially more compute per query than standard language models.

A custom inference chip optimised for OpenAI’s architectures can, in principle, cut those costs by 40 to 50 percent versus general-purpose NVIDIA GPUs.

At OpenAI’s scale, that’s a structural change in unit economics.

The pattern. Google has TPUs. Amazon has Trainium and Inferentia. Meta has MTIA. Microsoft is partnering on custom silicon. OpenAI completes the set.

NVIDIA’s training dominance isn’t under near-term threat, but inference, where production compute volume is consumed, is becoming a custom silicon market.

Why it matters

Every major AI company now has a path to inference independence from NVIDIA. That puts a ceiling on how much of the inference market NVIDIA can hold long-term.

Microsoft Is Spending $190 Billion On AI Infrastructure This Year. It Still Isn’t Enough.

Microsoft CFO Amy Hood disclosed this week that 2026 capex has been revised upward to $190 billion, against an average analyst forecast of $152 billion.

The $38 billion revision is driven mainly by surging prices for memory chips, storage, and GPU substrates, which Hood said accounted for $25 billion alone.

Despite spending at a pace unimaginable three years ago, Hood told investors Microsoft expects to remain capacity-constrained on GPUs, CPUs, and storage through at least the end of 2026.

The company currently has more contracted AI revenue than it can physically serve.

The revenue picture. Microsoft’s commercial remaining performance obligations have surged to roughly $625 billion. That’s contracted future revenue, not speculative pipeline.

Microsoft’s cloud and AI services are sold out. The constraint on revenue recognition isn’t demand. It’s physical infrastructure.

Every gigawatt of power Microsoft brings online converts directly into revenue it cannot currently book.

AI data centers require 80 to 140 kilowatts per rack versus 5 to 15 for traditional servers, meaning power, cooling, and construction costs are all significantly higher per unit of compute than historical norms.

Why it matters

Microsoft is spending $190 billion this year and still can’t serve all its contracted demand. That isn’t a bubble. It’s supply constrained by physical reality. The buildout is real, the demand is real, and the constraint is how fast the world can manufacture the components.

Google Redesigned Search For The First Time In 27 Years.

Google Search has operated on essentially the same interface paradigm since 1998: a text box, a button, a list of results.

At I/O this week, Google announced the end of that paradigm.

The new Search accepts text, images, files, videos, and open Chrome tabs as simultaneous inputs.

A user can share a screenshot, a PDF, a video clip, and a question all at once, and search reasons across them together rather than treating each as a separate query.

AI Overviews are now the default response format for complex queries, with traditional blue links appearing as supporting detail beneath.

Search Live. Announced alongside the redesign, it uses the device’s camera and microphone to answer questions about what the user is looking at or hearing in real time.

Point a camera at the equipment with an error message, and Search can explain it. Hold it up to a restaurant menu, and Search can recommend dishes based on dietary preferences.

The competitive significance. Google Search has roughly 90 percent of the global market share and processes about 8.5 billion queries per day.

If those queries shift from link-lists to AI answers, the economics of web publishing, advertising, and SEO all change at once.

Making AI Overviews the default is Google deciding the transition is happening whether it manages it or not, and that the risk of being displaced is greater than the risk of disrupting its own ad business.

Why it matters

Google just changed how Search works for the first time in nearly three decades. The text box is no longer the interface. AI answers aren’t a feature within Search. They are Search.

And that wraps up this week. Tune in next Monday, same time, for another deep-dive into the stories shaping the AI world.

The Sentinel lands in your inbox every Monday so you can catch up with the fast-moving AI space while sipping your morning coffee. Every detail that matters, none that doesn’t.

OpenAI launched a $4 billion deployment arm designed to live inside enterprise clients. Anthropic entered funding talks at a valuation above OpenAI’s. And Google disclosed the first confirmed AI-assisted zero-day prepared for mass attack.

Here’s everything you need to know before Monday gets the best of you.

OpenAI Stopped Acting Like A Software Vendor This Week.

On May 11, OpenAI launched the OpenAI Deployment Company, a majority-owned subsidiary backed by more than $4 billion from 19 investment firms, consultancies, and systems integrators.

TPG leads the partnership, alongside firms including Advent, Bain Capital, Brookfield, Goldman Sachs, SoftBank, Capgemini, and McKinsey.

The company is also acquiring AI consulting firm Tomoro, bringing roughly 150 Forward Deployed Engineers into the business immediately.

Those engineers work directly inside client organisations.

Not as support staff. Not as account managers. Their job is to rebuild workflows around AI systems and turn pilot projects into systems companies actually use every day.

Tomoro’s existing clients include Tesco, Virgin Atlantic, Red Bull, and Mattel.

This launch makes more sense when you look at what OpenAI has been struggling with internally.

Earlier this year, OpenAI chief revenue officer Denise Dresser reportedly told staff that Anthropic’s enterprise momentum should serve as a “wake-up call.” She also said OpenAI’s Microsoft exclusivity had limited the company’s ability to reach customers inside the infrastructure environments they already used.

The Deployment Company addresses both issues at once.

It gives OpenAI a direct enterprise channel that does not depend entirely on cloud partnerships, and it gives the company something most enterprise software firms spend years trying to build: embedded operational relationships.

The more revealing number came from OpenAI’s own State of Enterprise AI report published alongside the launch.

Despite roughly $40 billion in generative AI spending over the last two years, only 5 percent of enterprises report meaningful business returns.

That is the bottleneck OpenAI now appears focused on.

OpenAI seems to think the real problem is deployment, not the models themselves.

Embedding engineers inside organisations is slower and more expensive than selling API access. But it also tends to produce systems that companies actually rely on once the pilots end.

Why it matters

OpenAI is no longer behaving purely like a model provider. It is building a large-scale deployment business designed to sit inside enterprises directly. That puts it in competition with firms like Accenture, McKinsey, and the Big Four as much as other AI labs.

Anthropic Just Entered Funding Talks At A Valuation Above OpenAI’s.

Bloomberg reported this week that Anthropic is discussing a funding round of at least $30 billion at a valuation exceeding $900 billion.

The deal is still in early stages and has not formally closed, but the scale alone makes it one of the largest private technology raises ever attempted.

Expected participants reportedly include Greenoaks, Sequoia, Altimeter, and Dragoneer, with several firms considering commitments above $2 billion each.

If the round closes near current terms, Anthropic would surpass OpenAI’s March valuation and become the most valuable private AI company in the world.

The scale of the valuation sounds absurd until you look at the revenue curve underneath it.

Anthropic reportedly moved from roughly $9 billion in annualised revenue at the end of 2025 to more than $30 billion by April 2026.

Claude Code alone is said to be generating approximately $2.5 billion in annualised revenue.

Enterprise clients now make up the large majority of the business.

More than 1,000 enterprise customers reportedly spend over $1 million annually, and eight Fortune 10 companies are already clients.

Margins are improving too.

Gross margins reportedly climbed from 38 percent a year ago to above 70 percent, easing one of the biggest concerns investors previously had about frontier AI companies: whether compute costs would permanently crush profitability.

The capital itself has a fairly direct purpose.

Anthropic needs infrastructure.

The company is reportedly using the raise to secure enough compute capacity to remain competitive through a potential IPO that Bloomberg says could arrive as early as October.

The speed of the valuation jump is what stands out most.

Anthropic was reportedly valued at nearly $380 billion in February. Three months later, discussions are happening at more than double that number.

Why it matters

The market is no longer valuing frontier AI labs purely on future potential. Investors are now seeing revenue growth at a scale large enough to justify valuations that would have sounded impossible a year ago.

Google Found The First Confirmed AI-Assisted Zero-Day Being Prepared For Mass Attack.

Google’s Threat Intelligence Group published a report this week describing something security researchers have been anticipating for years.

A criminal actor used AI assistance to help identify and weaponise a zero-day exploit intended for a planned mass attack.

Google says the exploit targeted a widely used open-source web administration tool and was implemented in a Python script capable of bypassing two-factor authentication.

The company worked with the affected vendor before the attack was launched publicly and patched the vulnerability in advance.

Google did not identify the actor involved or the software affected.

The important part is how the exploit was apparently built.

According to Google’s assessment, the code shows strong signs of LLM assistance in how it was written and packaged.

This was not an autonomous system running attacks on its own.

A human threat actor still directed the process.

But AI clearly sped up the discovery and assembly of the exploit in a way that reduces the time and skill normally needed.

Defenders are already using AI for the same reason attackers are: speed.

Google’s broader report describes a growing ecosystem of AI-assisted cyber activity.

State-linked actors from China and North Korea are reportedly using AI for vulnerability discovery and malware development. Russia-linked groups have started embedding AI-generated obfuscation into Android malware workflows.

On the defensive side, Google’s Big Sleep agent is actively searching for unknown vulnerabilities, while CodeMender automatically patches certain categories of flaws.

The Five Eyes guidance on AI agents from last week reads differently after this report.

At the time, the warnings sounded precautionary.
Now they feel closer to documentation.

Why it matters

The security industry has spent two years debating whether AI would eventually assist real-world exploit development. This week produced the clearest confirmed example so far.

The EU Quietly Delayed Its Most Important AI Deadlines.

The European Parliament and Council reached a provisional agreement this week on amendments to the EU AI Act.

The biggest change is a 16-month delay to compliance deadlines for high-risk AI systems.

Rules covering systems used in areas like biometrics, critical infrastructure, education, employment, law enforcement, and border management are now expected to apply from December 2027 instead of August 2026.

Systems embedded in physical products move to August 2028.

The delay is mostly because companies still don’t have all the technical standards they need in order to comply properly.

Alongside the delay, the EU also added a new explicit prohibition on AI systems primarily designed to generate non-consensual intimate images, including so-called nudification apps and AI-generated child sexual abuse material.

That provision was pushed through by Parliament and was not part of the Commission’s original proposal.

Companies will have until December 2026 to comply.

The law itself is already in force.

What changes is when the hardest parts actually bite.

Smaller and mid-sized businesses also received broader exemptions designed to reduce compliance pressure, while registration requirements for certain lower-risk systems were tightened.

The EU is still building one of the world’s most comprehensive AI regulatory frameworks. It is just giving companies more time to implement it properly.

Why it matters

Europe didn’t step back from regulation this week. It just slowed the pace of enforcement on the hardest parts.

OpenAI’s Real-Time Voice APIs Make Always-On AI Assistants Technically Practical.

OpenAI released GPT-Realtime-2 and a new suite of voice APIs this week, focused on live audio interaction.

The difference from earlier voice systems is mostly latency and continuity.

Earlier systems converted speech into text, processed it, and then converted responses back into audio. The delay was small but noticeable.

GPT-Realtime-2 processes audio more directly, which cuts enough delay that conversations stop feeling like turn-based systems.

The APIs support real-time spoken conversation, live translation, and continuous transcription.

Uber is already testing the system inside its driver tools for ride booking and coaching workflows.

The bigger deal is that developers can now build all of this inside one system instead of stitching separate tools together.

And because the models preserve tone, pacing, and emphasis instead of flattening everything into text first, conversations feel more natural in situations where that matters.

Customer support. Coaching. Healthcare intake. Accessibility tools.

Those use cases become a lot more viable once the interaction stops feeling delayed.

At this point, the hard part is less technical and more about figuring out what people actually want to use every day.

Why it matters

Real-time conversational AI is moving out of demo territory and into production infrastructure. Voice interfaces just became much more viable across both consumer and enterprise software.

Only 5% Of Enterprises Report Real AI Returns. OpenAI Published The Number Itself.

OpenAI’s State of Enterprise AI report deserves attention separately from the Deployment Company launch.

The report surveyed more than 1,000 business leaders across six markets and documented a problem that much of the industry has been quietly running into for over a year.

Companies are spending heavily on AI.

Most are struggling to turn that spending into measurable business outcomes.

According to the report, enterprises have collectively spent close to $40 billion on generative AI initiatives over the last two years. Only 5 percent report meaningful returns.

Ninety percent of IT decision-makers also say they are reconsidering cloud strategies around AI workloads, balancing cost, performance, and control more aggressively than before.

The report argues that successful AI deployments tend to come from unified platforms with shared governance and production-grade architecture, rather than scattered pilot projects across departments.

Only 22 percent of organisations were classified as “future ready” from a data infrastructure perspective.

Most companies are trying to build AI systems on top of infrastructure that was never designed for it.

The timing of the report alongside the Deployment Company launch was deliberate.

OpenAI identified the problem and launched the proposed solution in the same week.

The sales pitch is fairly obvious: if 95 percent of enterprises are failing to generate returns, they become potential customers for a company promising to fix deployment itself.

Why it matters

The biggest obstacle to enterprise AI adoption increasingly looks like execution inside organisations, not model capability.

AI Infrastructure Just Hit Its Next Bottleneck. Not Chips. Electricity.

One of the quieter shifts this week came from the infrastructure side of the AI market.

The conversation around scaling AI used to be almost entirely about GPUs and chip supply. That is starting to change.

Bloomberg reported this week that Blackstone and Halliburton are backing VoltaGrid in a roughly $1 billion push to build gas-powered microgrids for AI data centres. The reason is simple: companies can secure chips faster than they can secure electricity.

That changes the economics of the industry.

Training clusters and large inference deployments are now running into grid capacity limits, utility approval delays, and energy pricing pressure before they hit compute constraints.

Several major infrastructure projects this year already include dedicated energy agreements tied directly to data centre expansion.

Power is starting to behave as GPUs did in 2024: something you have to secure early if you want to scale.

That has ripple effects.

It influences where data centres get built, how fast cloud providers can expand capacity, and which regions can realistically compete for large-scale AI infrastructure.

Two years ago the constraint was GPUs. Now it is starting to look more like electricity and physical infrastructure.

Why it matters

The limiting factor for scaling AI is shifting beyond chips and toward energy itself. Whoever can secure reliable power capacity ends up with a structural advantage in the next phase of the industry.

And that wraps up this week. Tune in next Monday, same time, for another deep-dive into the AI world.

The Sentinel lands every Monday, so you can catch up without having to sit through the noise.

Here’s what happened in the global AI market this week.

AI beat emergency room doctors on real patients in a study published in Science. Six governments issued the first international security guidance focused specifically on AI agents. OpenAI rolled out a clinician-grade AI tool to verified medical professionals across the United States. And at NVIDIA’s biggest conference of the year, enterprise teams spent more time discussing governance and reliability than model benchmarks.

Here’s everything you need to know before Monday gets the best of you.

AI Beat Real ER Doctors On Real Patients. The Study Got Published In Science.

Researchers from Harvard Medical School and Beth Israel Deaconess Medical Center published the most serious real-world comparison yet between frontier AI models and working physicians.

The paper landed in Science this week.

Researchers tested OpenAI’s o1 model against doctors across six separate clinical experiments. In every one, the model performed on par with or better than the physicians involved.

The most important test used 76 real emergency room patients from Beth Israel. No cleaned-up benchmark data. No idealised examples. Same records, same information, same time pressure.

At triage, the model identified the correct or near-correct diagnosis 67 percent of the time. The physicians scored 55 percent and 50 percent. As more information became available during admission, everyone improved. The model reached 81.6 percent. The doctors reached 78.9 percent and 69.7 percent.

The structure of the study matters here. Two attending physicians issued diagnoses. So did o1 and GPT-4o. A separate pair of attendings then graded every answer blind, without knowing which responses came from humans and which came from models.

The gap held consistently.

On clinical reasoning, which measured how well diagnoses were explained and what next steps were recommended, o1 received a perfect score on 98 percent of cases. The attending physicians hit 35 percent. On treatment planning across five clinical case studies, the model scored 89 percent. Forty-six doctors using conventional resources, including search engines, scored 34 percent.

Lead author Adam Rodman said the most surprising result was that the model handled messy emergency room data reliably. That is different from performing well on curated benchmark cases.

The caveats are still important. The model only had access to text. Real clinicians see the patient, hear breathing, read body language, and make decisions using information that never enters the medical record. Emergency physician Kristen Panthagani also pushed back publicly on the framing of some headlines, noting the comparison group used internal medicine physicians rather than ER specialists.

But the broader point is difficult to ignore.

AI systems are now performing at a level where hospitals, insurers, and regulators all have the same problem to solve: responsibility.

Why it matters

This study changes the healthcare AI discussion from capability to accountability. The question is no longer whether models can assist with clinical reasoning. It is who takes responsibility when those systems are used in real care settings.

Six Governments Just Issued The First Security Warning Written Specifically For AI Agents.

Six national security agencies published a joint warning about AI agents this week.

A year ago, most cybersecurity discussions around AI were still focused on chatbots, phishing, and synthetic content. This document is about autonomous systems operating inside real infrastructure.

On May 1, CISA, the NSA, Australia’s ASD, the Canadian Centre for Cyber Security, New Zealand’s NCSC, and the UK’s NCSC released a 30-page document titled “Careful Adoption of Agentic AI Services.” It is the first coordinated international security guidance written specifically for AI agents rather than AI systems broadly.

The agencies identified 23 separate risk categories and more than 100 recommended practices. Three issues received the most attention.

First: privilege escalation. Agents are routinely being granted broader system access than their tasks require, which means a compromised tool inside the workflow can inherit those permissions.

Second: prompt injection. Malicious instructions hidden inside documents, emails, or webpages can manipulate agents into taking actions operators never intended.

Third: accountability gaps. Multi-agent systems hand tasks across multiple models and tools until reconstructing the chain of decisions afterward becomes extremely difficult.

The Register described one possible scenario plainly: compromise a low-risk tool connected to an agent, inherit the agent’s permissions, approve fraudulent payments, modify contracts, then generate fake audit logs afterward.

The guidance was prompted by observed incidents.

Most of the recommendations are surprisingly basic. Least-privilege access. Human approval for irreversible actions. Independent audit logging. Treating agents as untrusted systems when they interact with each other.

The difficulty is not the advice itself. It is that organisations are deploying agents faster than they are building the processes needed to supervise them properly.

Why it matters

When six intelligence agencies coordinate on guidance for a specific technology category, it usually means the technology has already moved beyond experimentation and into operational risk.

OpenAI Put A Clinician-Grade AI Tool In Front Of Every Verified Doctor In America.

The rollout landed in the same week as the Harvard study.

On April 23, OpenAI launched ChatGPT for Clinicians, a version of ChatGPT designed for clinical workflows and made available free to verified physicians, nurse practitioners, physician assistants, and pharmacists in the United States.

The product focuses on documentation, medical research, and care coordination.

Doctors can use it to draft notes, summarise literature, prepare referral letters, generate discharge summaries, and assist with multi-step clinical reasoning. Before release, OpenAI says physician advisors tested 6,924 real-world conversations inside clinical settings. Doctors rated 99.6 percent of responses as safe and accurate.

Alongside the launch, OpenAI also published HealthBench Professional, an open benchmark designed to evaluate medical AI systems using physician-written conversations and specialty-specific grading rubrics.

The benchmark result is the important part.

GPT-5.4 inside the Clinicians workspace outperformed every other tested system, including the human physician baseline. And the physician baseline was not weak competition. Specialty-matched doctors were given unlimited time and web access.

Healthcare AI has been discussed for years as a future possibility. The combination of this rollout and the Harvard study makes it look much more immediate.

The unresolved issue is responsibility. If an AI-assisted clinical decision harms a patient, who carries liability for that outcome? The hospital? The physician? The vendor? The answer is still unclear.

The deployment is happening before those questions are settled.

Why it matters

This was not a limited pilot program. OpenAI rolled out a nationwide clinical AI product while the legal and regulatory framework around clinical AI is still being worked out.

The US Government Quietly Moved Closer To Reviewing Frontier AI Models Before Release.

Governments are starting to treat frontier AI systems less like ordinary software products and more like infrastructure with national security implications.

Multiple reports this week confirmed that companies including Google, Microsoft, and xAI are allowing advanced AI systems to be evaluated by US government agencies before broader public deployment. The reviews are reportedly focused on cyber capabilities, infrastructure exposure, autonomous behaviour, and misuse risks.

The systems involved are not ordinary chatbots. Agencies are reportedly evaluating models capable of reasoning, writing code, interacting with tools, and operating inside sensitive environments.

Until recently, the idea of frontier labs voluntarily giving governments early access to advanced systems before release would have been controversial across much of the industry. Now it is happening quietly and with relatively little resistance.

The timing also matters. The reports landed the same week six allied intelligence agencies issued the first coordinated guidance focused specifically on AI agents. Together, the developments suggest governments are moving past broad discussions about AI risk and toward direct oversight of specific capabilities.

The process still appears cooperative rather than mandatory. But once governments become accustomed to reviewing frontier systems before deployment, formal oversight becomes much easier to justify politically.

Why it matters

For the past two years, most AI regulation discussions have been theoretical. Governments are now starting to involve themselves much earlier in the deployment process for frontier systems.

JPMorgan Stopped Calling AI A Technology Investment. Now It Calls It Infrastructure.

JPMorgan’s latest technology budget revealed something more important than the spending itself.

AI is now classified alongside cybersecurity and operational resilience.

In a widely circulated interview this week, JPMorgan CIO Lori Beer said the bank’s 2026 technology budget will reach $19.8 billion, up 10 percent year-on-year. Roughly $2 billion is directly tied to AI initiatives.

The bigger signal was where that spending got classified.

AI is no longer being treated as a discretionary innovation project inside the bank. It now sits inside the same category as the systems required to keep the institution operational.

That distinction matters because infrastructure budgets survive economic downturns, executive turnover, and shareholder pressure in ways experimental spending usually does not.

JPMorgan now has more than 2,000 employees dedicated to AI systems, which scan roughly $10 trillion in daily transactions for fraud, anomalies, and optimisation opportunities. The bank says realised annual AI value already roughly matches annual AI spend.

The systems are already paying for themselves.

Beer also said AI deployments inside the bank doubled in 2025, with generative AI becoming the fastest-growing category. Most of the gains are showing up in software engineering and customer-facing systems, where development cycles are shrinking and response times are improving.

The important part is not that JPMorgan is spending aggressively on AI. Most large banks are doing that.

It is that the bank no longer appears to view the spending as temporary.

Why it matters

When the largest bank in the United States reorganises its technology budget to treat AI as part of core infrastructure, the rest of the financial industry pays attention.

NVIDIA’s Biggest Conference Of The Year Barely Talked About Chips.

For years, NVIDIA’s GTC conference revolved around hardware performance, benchmarks, and model capability.

This year the tone was noticeably different.

Jensen Huang used the keynote to describe AI as a core operating layer for industry, but the sessions drawing the largest crowds focused less on raw capability and more on deployment reliability.

Manufacturing companies. Logistics firms. Financial institutions.

The engineers building these systems were asking operational questions now. How do you monitor autonomous systems properly? How do you audit them? How do you enforce security controls when agents are making decisions independently?

NVIDIA’s own announcements reflected the shift. The company launched its Agent Toolkit for enterprise AI agents alongside OpenShell, a runtime designed to enforce security and policy controls at the infrastructure layer instead of relying entirely on model behaviour.

The AI-Q architecture also drew attention by combining frontier orchestration models with open Nemotron systems for retrieval and research tasks, reducing enterprise query costs significantly while maintaining strong performance.

But the most revealing part of GTC was probably the mood around the conference itself.

The engineers building enterprise AI systems were spending less time debating whether the models were capable enough and more time discussing what happens once those systems are deployed at scale.

You could hear it in the questions people were asking.

Why it matters

The biggest enterprise AI conference in the world spent more time discussing governance, observability, and reliability than benchmark performance. That says a lot about where the industry thinks the difficult problems are now.

Caltech Built An AI That Segments Every Cell In Every Image. Biologists Just Got Their Time Back.

A surprising amount of biological research still begins with manual image annotation.

Researchers published CellSAM in Nature Methods this week to reduce that bottleneck.

The model handles biological image segmentation, the process of outlining individual cell boundaries, before researchers can analyse microscopy images properly. The task is repetitive, time-consuming, and often takes hours before the actual scientific work even begins.

CellSAM automates that process across multiple cell types, imaging conditions, and biological contexts.

Previous segmentation systems usually performed well only inside the narrow environments they were trained on. Models trained on one tissue type often struggled badly when applied elsewhere, forcing researchers to retrain systems constantly or correct outputs manually afterward.

CellSAM appears to generalise much more reliably across real laboratory conditions.

A graduate student who previously spent two hours annotating cells can now validate outputs in minutes and move on to the actual analysis.

That changes research workflows in a meaningful way.

The model also runs on standard laboratory hardware rather than requiring expensive cloud infrastructure, which makes adoption easier for smaller research labs without large computing budgets.

Cell segmentation sits upstream of a huge amount of biological research. If that step becomes dramatically faster, study sizes increase, timelines compress, and experiments that were previously too labour-intensive become more practical.

The researchers who built the model described the motivation pretty simply: too much scientific time was being spent on work computers should already be able to handle.

Why it matters

Reliable automated segmentation removes one of the most time-consuming preparation steps in modern biological imaging workflows. That has downstream effects on how quickly labs can run and analyse experiments.

And that wraps up this week. Tune in next Monday, same time, for another deep-dive into the stories shaping the AI world.

The Sentinel lands in your inbox every Monday so you can catch up with the fast-moving AI space while sipping your morning coffee. Every detail that matters, none that doesn’t.

Here’s what happened in the global AI market this week.

Microsoft’s exclusive grip on OpenAI ended. Google put $40 billion behind its biggest AI rival. Coding agents got the ability to buy domains, provision servers, and ship software without asking anyone. And the trial that could rewrite how AI companies are allowed to exist finally opened in Oakland.

Here’s everything you need to know before Monday gets the best of you.

Microsoft’s Hold On OpenAI Is Over, And The Cloud War Just Opened Up

On April 27, Microsoft and OpenAI announced they had torn up the exclusivity clause that had bound OpenAI’s commercial products to Azure since before ChatGPT launched. The renegotiated deal gives Microsoft a non-exclusive licence to OpenAI’s IP through 2032, caps Microsoft’s revenue share, and eliminates the AGI clause that would have changed their commercial relationship the moment artificial general intelligence was declared. OpenAI is now free to run its products on any cloud. Within 24 hours, it was on Amazon Web Services.

AWS CEO Matt Garman held an event in San Francisco on April 28 to launch OpenAI models on Amazon Bedrock, framing it as meeting years of suppressed demand. “Their production applications run in AWS. Their data is in AWS. We’ve forced them to get great OpenAI models, to go to other places.” That forced detour is over. OpenAI CEO Sam Altman appeared at the event via recorded video, noting that his schedule had been “taken away” from him for the day, a reference to the Musk trial opening across the Bay Bridge in Oakland. The new product is called Amazon Bedrock Managed Agents, powered by OpenAI, combining OpenAI’s models with AWS’s Trainium infrastructure, with OpenAI committing roughly 2 gigawatts of Trainium capacity for training. Google Cloud is already in advanced talks for the same arrangement, targeting Q4 2026. GPT-5.5, which shipped on April 23 and which we talked about on The Sentinel last week, is already available on Bedrock, alongside Codex and the full model suite.

The internal OpenAI memo that surfaced this week, written by chief revenue officer Denise Dresser, is the most revealing part of the story. She wrote that the Microsoft partnership had “limited our ability to meet enterprises where they are” and that inbound demand for the AWS offering had been “frankly staggering.” That’s harsh language for a company that willingly stayed exclusive for three years. The deal was a constraint. Removing it unlocks enterprise customers who spent the past year asking why they couldn’t get OpenAI models inside the infrastructure they already run. Well, now they can. And AWS, which already carries Anthropic’s Claude and Meta’s Llama family, has quietly become the most comprehensive AI model marketplace in existence.

Why it matters

Every enterprise that had Azure as its primary reason to avoid OpenAI’s models lost that excuse this week. AWS now carries OpenAI and Anthropic under the same roof. That is a different competitive landscape than the one that existed seven days ago.

Google Is Betting $40 Billion On Anthropic While Also Competing With It

Google confirmed on April 24 that it is committing up to $40 billion to Anthropic, starting with a $10 billion investment at a $380 billion valuation, with the remaining $30 billion contingent on performance milestones. This is not a new relationship. Google has been invested in Anthropic since its earliest funding rounds. But the scale is different this time. The deal pairs the cash commitment with a fresh five gigawatts of computing capacity from Google Cloud over five years, adding to a previously announced 3.5 gigawatt TPU arrangement with Google and Broadcom from 2027. Anthropic now has more committed compute than any other frontier lab outside of the hyperscalers themselves.

The $10 billion lands immediately. The other $30 billion arrives only if Anthropic hits specific operational and financial milestones Google has not publicly named. That structure matters. At a $380 billion valuation, getting $30 billion conditional rather than committed reflects the risk of investing in a company that is also competing directly against your core search and AI products. Google needs Anthropic to be competitive enough to stay at the frontier, but not so dominant that Claude makes Google’s own AI efforts irrelevant. The five-gigawatt compute commitment is the part that solves a concrete problem: Anthropic can now train the next two generations of models without infrastructure problems.

OpenAI went multi-cloud the same week, with its models landing on AWS and Google Cloud talks already underway. A search giant that hosts both of its main AI competitors while also investing $40 billion in one of them. This game is far more complex than it appears. Google’s bet is that controlling the infrastructure layer is worth more than backing any single model. If AWS becomes the distribution platform for every frontier model, Google loses, regardless of how good Gemini is. Keeping Anthropic well-capitalised and compute-secure is part of how Google stays relevant in the infrastructure race it cannot afford to lose.

Why it matters

Five gigawatts of compute paired with $40 billion in capital is the largest infrastructure commitment any investor has made to a frontier AI lab. Anthropic won’t be struggling for compute for a long time. That changes what it can build next.

AI Agents Can Now Pay, Deploy, And Ship Without Asking Anyone

On April 30, Cloudflare published a protocol co-designed with Stripe that allows AI coding agents to provision cloud accounts, start paid subscriptions, register domain names, and deploy applications to production, without a human touching a dashboard or entering a credit card number at any point. Three steps: discovery, where the agent queries a service catalog; authorisation, where Stripe verifies the user’s identity and Cloudflare provisions an account automatically; and payment, where Stripe passes a tokenised credential so the agent can pay for services without ever handling raw card details. Stripe sets a default spending cap of $100 per month per provider. Nothing to a live app on a registered domain in just a few minutes.

It shipped on April 30 via the Stripe Projects CLI, currently in open beta, with Cloudflare as the first infrastructure partner. Additional partners already live include Vercel, Supabase, Hugging Face, Twilio, and AgentMail. The protocol is open. Any platform with signed-in users can play the role Stripe does, acting as the identity provider for any agent-accessible service. The security concerns are also there and in consideration. David Shipley of Beauceron Security pointed out that this also dramatically accelerates the ability of bad actors to spin up infrastructure, noting that cyber criminals already need to constantly create new accounts as security firms shut old ones down. Faster provisioning helps everyone. Including the people you don’t want it to help.

An agent that writes code, tests it, provisions the infrastructure, buys the domain, and ships the app is not a developer’s assistant. It is a system that builds and ships software. Who owns the subscriptions it opened? Who is liable for the spend? How do you revoke access across a dozen provisioned services when something goes wrong? Those questions don’t have clean answers yet. Enterprise teams will be working through them for the rest of 2026. But the pipeline itself, from idea to live product, is now fully automated and available to anyone with a Stripe account.

Why it matters

The last step that required a human in an autonomous coding workflow is gone. Not coming soon. Gone. That happened on April 30.

Musk’s Lawsuit Isn’t About The Money. It’s About Setting The Rules

The trial began on April 28 in the Northern District of California, Oakland. Elon Musk is suing OpenAI and Sam Altman for over $150 billion in damages, arguing that OpenAI defrauded him as a founding donor by abandoning the nonprofit mission under which it was established, then converting to a for-profit structure without honouring the commitments it made when it accepted his money. Altman was across the Bay Bridge at the AWS event in San Francisco, attending via recorded video. “My schedule got taken away from me today,” he said, which landed somewhere between diplomatic and pointed depending on who you asked.

What is not in dispute is that Musk donated approximately $45 million to OpenAI between 2015 and 2017, left the board in 2018, and OpenAI is now mid-conversion to a for-profit public benefit corporation. What is very much in dispute: whether Altman and the board made specific representations about the organisation’s permanent nonprofit status, whether those representations were legally enforceable, and whether the for-profit conversion constitutes fraud. OpenAI’s defence is that the mission was always to develop safe AI for humanity’s benefit, not to remain a nonprofit structure indefinitely, and that the conversion serves rather than betrays that mission.

The amendment Musk filed before trial changes everything about how to read this case. He is now asking that any damages be awarded to OpenAI’s charitable arm rather than to himself, and that Altman be removed from the nonprofit board. The personal financial gain angle is gone. What remains is a governance fight over whether the most prominent AI organisation in the world can be held to its founding commitments. OpenAI’s for-profit conversion is not yet complete. A ruling against them could halt it. A ruling for them sets a precedent that AI labs can walk away from their founding missions without legal liability to the donors who funded them. That precedent, not the $150 billion, is what every lab in the industry is watching from Oakland.

Why it matters

This case will decide whether AI organisations owe anything to the people who funded their early work. The answer becomes precedent for every lab currently navigating pressure to restructure. The outcome in Oakland matters far beyond OpenAI.

Salesforce Is Rebuilding Itself For AI Agents, Not Human Users

Salesforce announced Headless 360 at its TDX developer conference in San Francisco, describing it as the biggest architectural change in the company’s 27-year history. Everything on Salesforce, every CRM record, every workflow, every business logic rule, every approval process, is now accessible as an API, an MCP tool, or a CLI command. An AI agent can read a customer support queue, pull the case history, trigger the resolution workflow, update the record, and close the ticket without a human opening a browser. Co-founder Parker Harris put it from the stage: “Why should you ever log into Salesforce again?” Not a rhetorical question. An architectural statement.

The release ships more than 60 new MCP tools and 30 preconfigured coding skills, giving external coding agents, Claude Code, Cursor, Codex, and Windsurf complete live access to a customer’s entire Salesforce org. Agentforce Vibes 2.0, the company’s native development environment, now supports both the Anthropic agent SDK and the OpenAI agents SDK. The pricing model is shifting from per-seat to consumption-based, which is a quietly significant part. Per-seat pricing assumes a human for every seat. Consumption-based pricing assumes work is the unit, not the worker. Salesforce is betting the number of human seats is going in one direction, and building its revenue model around what comes after.

SaaStr ran a piece this week from a company that has been running this architecture for six months already, three humans and twenty-plus AI agents in production, and almost nobody logging into the Salesforce UI. “Going AI-first didn’t make CRM less important,” they wrote. “It made us finally use it the way it was always supposed to be used: as the data backbone, with intelligence running on top.” That is the product Salesforce is now selling. Not software for humans to click through. Infrastructure for agents to operate.

Why it matters

Salesforce has $30 billion in annual revenue built on the assumption that humans use software through graphical interfaces. Headless 360 is the company betting that assumption has an expiry date, and choosing to get ahead of it.

Mistral Isn’t Chasing Better Models. It’s Fixing Why AI Fails In Production

Mistral released Mistral Workflows this week, a production orchestration engine for multi-step AI pipelines. The target is enterprise teams that have built successful AI experiments and then hit a wall trying to put them into production. The product handles the parts that kill deployment: state management across multiple steps, tool call sequencing, error handling and retry logic, full observability into what the pipeline is doing at each stage, and privacy controls over which data each step can access. Not capabilities for making AI more impressive. Infrastructure for making it reliable enough to trust in production.

The observability piece is where teams consistently get stuck. They cannot see inside what the system is doing well enough to debug failures, satisfy audit requirements, or make a credible case to the legal and compliance teams that need to approve anything running in production. Mistral Workflows ships with built-in logging, step-by-step trace views, and integration with standard enterprise monitoring tooling. The privacy controls, defining what data each pipeline step can touch, are aimed at the same audience. These are features that go way beyond just benchmark numbers to actually getting deals signed.

Mistral has been building its enterprise business quietly while OpenAI and Anthropic take the headlines. Its models run cheaper than frontier alternatives. Its Paris headquarters gives it a compliance advantage for organisations under GDPR and EU AI Act obligations that its American competitors cannot replicate without a structural change. Workflows is not a play for the “which model is most capable” conversation. It is a play for the deployment layer underneath that conversation, the infrastructure that turns a capable model into something an enterprise actually runs. The strategy is to let the big labs fight over who has the best model, and own the layer that makes any model usable.

Why it matters

Most enterprise AI projects stall on deployment, rather than capability. Mistral shipped the tooling to fix that problem, with compliance features baked in from the start rather than added later. In regulated industries, that matters more than any other benchmark point.

China Is Winning The In-Car AI Race. The West Is Far Behind

ByteDance confirmed this week that its Doubao AI assistant is now integrated into 145 electric vehicle models across China, making it one of the most widely deployed in-vehicle AI systems in the world by model count. The integration goes well beyond voice commands and navigation. Doubao handles real-time contextual awareness of the journey, personalised content recommendations during transit, natural language interaction with vehicle systems, and multi-turn conversations that persist across sessions. Alibaba’s Qwen and Baidu’s ERNIE are fighting for the same cockpit real estate. The Chinese EV market has become the most competitive deployment environment for conversational AI in any physical product category, anywhere in the world.

China’s EV market shifted from competing on range and charging speed to competing on the intelligence of the in-vehicle experience roughly 18 months ago. BYD, NIO, Li Auto, and Xpeng have all treated cockpit AI as a primary product differentiator. ByteDance’s advantage is not its model quality. It’s its data. Douyin has over 700 million daily active users. Doubao was the fastest AI app to reach 100 million users in China. The personalisation dataset that it generates gives it an edge that purpose-built automotive AI companies cannot replicate without years of consumer-facing deployment at scale. The 145-model figure is the advantage being converted into automotive partnerships.

Chinese EV manufacturers are expanding into Europe, Southeast Asia, and Latin America. The in-vehicle AI that dominates the Chinese market travels with those vehicles. Doubao, Qwen, and ERNIE are not household names in Western markets yet. In two years, they could be the AI that hundreds of thousands of drivers interact with for an hour every day, a higher-frequency touchpoint than most people have with any AI system. The battle for the cockpit is being fought almost entirely in China. Western AI labs are not seriously competing for it. They may want to reconsider that.

Why it matters

The car cockpit is the most overlooked AI deployment surface in the world. Chinese labs are winning it. As Chinese EVs go global, so does their AI. That’s a distribution story the Western AI industry hasn’t priced in yet and probably should.

And that wraps up this week. Tune in next Monday, same time, for another deep-dive into the stories shaping the AI world.

The Sentinel lands in your inbox every Monday so you can catch up with the fast-moving AI space while sipping your morning coffee. Every detail that matters, none that doesn’t.

Here’s what happened in the global AI market this week.

Anthropic unveiled a 10 trillion-parameter model and kept it restricted. DeepSeek V4 landed and exposed the reality of the chip race. Meta is building its own silicon stack. And an AI system just passed peer review on its own.

Here’s everything you need to know before Monday gets the best of you.

Anthropic Built a 10T Model, But Not Everyone Gets Access

Anthropic this week unveiled Claude Mythos 5, its most powerful model to date and the first publicly announced AI system to cross the 10 trillion parameter threshold. For context: GPT-4 was estimated at around 1.7 trillion. In roughly two years, the scale has increased six times over. The model introduces a new tier in Anthropic’s lineup, codenamed Capybara, sitting above Opus, and is being positioned for cybersecurity analysis, advanced scientific research, and complex software engineering. Not a general-purpose upgrade. A specialist instrument for high-stakes work.

Anthropic has not published official benchmarks and the model is not publicly available. Access is restricted to a controlled preview through Project Glasswing, the same coordinated security initiative that emerged from the zero-day vulnerability discoveries covered 2 weeks prior. The list of approved organisations includes AWS, Apple, Cisco, Google, JPMorgan Chase, Microsoft, Nvidia, CrowdStrike, and around 40 other critical infrastructure partners. All access is limited to defensive cybersecurity work. Anthropic described the model as “far ahead of any other AI model in cyber capabilities” in internal documents that were briefly and accidentally made public in late March when a content management misconfiguration exposed roughly 3,000 internal assets. The company locked them down within hours, but the damage had been done.

The 10 trillion figure does not automatically mean 10 times better at everything. The relationship between parameter count and capability is logarithmic, not linear. What it does mean is that Anthropic has access to a model capable of sustained reasoning over extremely long contexts, multi-step problem decomposition, and domain depth that smaller models can’t match. The training run is estimated to have cost several hundred million dollars. No API pricing has been published. And with GPT-6 still unshipped, Mythos 5 is the most capable model any lab has publicly confirmed this week.

Why it matters

Anthropic built something so capable it decided the public shouldn’t have unrestricted access to it. That’s a signal about what this category of model can actually do. Every other lab is now working out what their equivalent looks like.

DeepSeek V4 Finally Lands And Proves the AI Race Isn’t Slowing Down

DeepSeek released V4 on April 24 after missing three projected windows. The wait is over. Two models shipped simultaneously: V4-Pro, with 1.6 trillion total parameters and 49 billion active per response, and V4-Flash, a leaner 284 billion total parameter version built for speed and cost. V4-Pro is now the largest open-weight model in existence, bigger than Kimi K2.6 at 1.1 trillion and more than twice the size of DeepSeek’s own V3.2. Both are available on Hugging Face under an MIT licence, free to download, run, and modify. The API went live the same day at $1.74 per million input tokens and $3.48 per million output for Pro, a fraction of GPT-5.5’s $5 and $30 and Opus 4.7’s $5 and $25. DeepSeek’s pricing has historically reshaped what the rest of the market can charge. V4 is no exception.

The architecture running underneath is worth understanding. Both models use a Hybrid Attention Architecture, DeepSeek’s term for a design that dramatically improves memory efficiency at extreme context lengths. V4-Pro at 1 million tokens uses only 27 percent of the single-token inference FLOPs and 10 percent of the KV cache size of V3.2. That is not a minor optimisation. It means genuinely usable 1 million token context at a cost that makes it viable for production, not just benchmarks. On formal mathematics, V4 stretches past the field: Putnam-2025 proof verification at a perfect 120 out of 120, and Codeforces at a 3,206 rating, ranking 23rd among human competitive programmers. On standard agentic coding and reasoning benchmarks it sits between GPT-5.2 and GPT-5.4, which means GPT-5.5, released one day earlier, pushed the closed frontier further out. The honest reading: V4 is not the strongest model in the world. It is the strongest model anyone in the world can freely download and run themselves.

The Huawei story matters too. V4 was built on domestic Chinese silicon. The three months of delays were not about the model being unready. They were about Huawei’s Ascend 950PR chips not producing units fast enough to run the scale of testing DeepSeek needed. The model shipped when the hardware supply caught up. A frontier-competitive open-source model, trained and deployed entirely on chips with no American component, now exists. MIT Technology Review called it “an early sign that China is successfully building a parallel AI infrastructure.” That framing is accurate and understated.

Why it matters

The most capable open-source model ever built is now free to download. It was trained on Chinese silicon, costs a fraction of every closed competitor to run via API, and just proved that export controls are not the ceiling anyone hoped they were.

Meta Is Building Its Own Chips And Redefining AI Infrastructure

Meta and Broadcom announced an expanded partnership to co-develop four generations of Meta’s custom MTIA chips, the Meta Training and Inference Accelerator, through 2029. The initial deployment commitment exceeds 1 gigawatt of compute. That is the first phase of a sustained multi-gigawatt rollout. Broadcom CEO Hock Tan stepped off Meta’s board the same day the deal was announced, given the scale of the commercial tie-up. Mark Zuckerberg’s own words on the announcement: “build out the massive computing foundation we need to deliver personal superintelligence to billions of people.”

The architecture matters. MTIA is Meta’s purpose-built accelerator optimised for inference and recommendation at scale that runs every time someone opens Instagram, scrolls Facebook, or asks Meta AI something. Four new generations of the chip are being developed and deployed within two years. The first to hit 2nm process technology will be a major performance leap. Broadcom’s XPU platform gives Meta the design and packaging capability to build silicon tuned to its own workloads rather than relying on general-purpose GPU architectures that aren’t optimised for what Meta actually runs at scale.

Meta joins Google, Amazon, and Microsoft in pursuing custom accelerators alongside continued Nvidia spend. The pattern is now stable across every major hyperscaler. Use Nvidia for leading training capacity, use custom chips for the inference and recommendation workloads that run continuously at enormous volume. Nvidia still gets paid. But the share of infrastructure spend flowing to custom silicon is growing with every generation. The long-term trajectory is clear.

Why it matters

Every hyperscaler building its own chips is a vote of no-confidence in Nvidia’s long-term monopoly on AI compute. Meta’s MTIA roadmap isn’t a hedge. It’s an infrastructure declaration.

Spotlight

Krater Just Made Your AI Work While You Sleep

Most AI tools stop working the moment you close the tab. Krater’s two new features change that.

Additions is a library of add-ons with three types: Personas, Prompts, and Apps. Apps are the interesting ones. One click connects Gmail, Google Calendar, Notion, Slack, GitHub, Linear, or hundreds more. Once connected, Krater can read from and act on those services directly inside any chat. “Summarize my unread Gmail.” “Find conflicts in my calendar next week.” “Create a Linear ticket from this conversation.” Connected apps stay linked to the account and are pinned to the top of the library so they’re always one tap away.

Tasks is the scheduler. Any prompt worth repeating, a morning briefing, a weekly competitor scan, a daily inbox summary, can be turned into a recurring task with a name, a model, a cadence, and delivery preferences. Results land in a dedicated chat, as a notification, or straight to email. The real unlock is that Tasks run through the same engine as regular chats, which means they inherit every connected app. That means the task isn’t limited to what the model knows. It reaches into the user’s actual tools. Set “Every weekday at 8 am, check my Gmail for anything urgent, cross-reference my calendar, and tell me what to focus on,” and Krater does exactly that, every morning, before the user sits down. Apps give Krater hands. Tasks give it a clock. Put them together, and the AI starts working for you while you sleep.

Try it

Additions and Tasks are live on Krater now. If your AI has been waiting for you to show up, this flips that around.

Meta And Microsoft Fire 20,000 Employees As Big Tech Doubles Down on AI Spend

Meta and Microsoft announced major workforce reductions on April 23 and 24. Meta is cutting 10 percent of its global workforce, around 8,000 employees, with cuts beginning May 20. Microsoft introduced a voluntary separation programme, its first in 51 years, affecting another 12,000 to 15,000 positions. Together, more than 20,000 jobs across two of the largest technology companies in the world are being eliminated in the same week. The stated rationale from both companies is nearly identical: AI is enabling smaller teams to do the same work, and the capital freed up from headcount is being redirected into infrastructure and AI development. The cumulative number of tech workers laid off in 2026 now exceeds 92,000.

The number worth sitting with is not the headcount. It is the capital being deployed in the other direction. Alphabet, Microsoft, Meta, and Amazon are collectively spending close to $700 billion on AI infrastructure this year. Meta alone has committed up to $135 billion on AI capital expenditure in 2026. The same companies cutting staff are pouring more money into computing, data centres, and AI development than any group of companies has spent on a single technology in any single year in history. The workforce reductions are not a sign of struggle. They are a sign that the output-per-human ratio is shifting fast enough that maintaining 2024 headcount levels looks like overstaffing to their boards.

The harder question is whether AI is actually causing this or whether companies are using it as cover. A reasonable reading suggests both are true at once. Meta is closing its Horizon World metaverse platform by June, which alone accounts for thousands of roles in Reality Labs. Microsoft’s buyout programme looks more like a skills transition than pure technology displacement. But Block invoked AI, cut 4000 jobs, and its stock rose the following day. Snap cut 1000 and rose 7 percent. Meta cut 8000, and the market rewarded it. Whatever the underlying cause, markets have now clearly signalled that AI-attributed headcount reductions increase valuation. That incentive structure is not going away.

Why it matters

Twenty thousand jobs at two of the world’s most valuable companies in one week, with $700 billion being spent on AI in the same year by the same companies. This is what the transition looks like in financial statements. It is no longer coming. It is the current quarter.

An AI System Published A Research Article And Passed Peer Review Without Human Authors

Sakana AI’s AI Scientist v2 generated the first fully AI-authored paper to pass a rigorous human peer-review process, with the achievement now published and documented in Nature. The system is end-to-end. It generates hypotheses, designs and runs experiments, analyses data, visualises results, and writes the manuscript in LaTeX with citations. Sakana submitted three fully autonomous papers to a peer-reviewed ICLR workshop. One exceeded the average human acceptance threshold. That’s the one that made history.

The v2 system is a meaningful architectural step beyond its predecessor. Where v1 relied on human-authored code templates to frame the research, v2 eliminates that requirement and generalises across machine learning domains. It uses a progressive agentic tree search, branching into multiple parallel experiments, picking the most promising results, iterating, and backtracking when it hits dead ends. A dedicated experiment manager agent coordinates the process. A vision-language model checks figures for clarity and accuracy. The whole thing costs around $15 per paper in API and compute fees. Sakana acknowledges that the more complex v2 search raises that figure somewhat. But even at 10x the cost, it remains a fraction of what human research teams spend.

The system has its limitations. It doesn’t produce better papers than v1 all the time. Its literature review process has been criticised for relying on keyword searches rather than proper synthesis. The quality of the manuscripts has been compared to an undergraduate rushing a deadline. But that comparison is the point. An undergraduate can be educated. An AI system that can already pass peer review can be iterated on. The current quality ceiling isn’t that high, but the trajectory is spot on.

Why it matters

The first peer-review milestone has been crossed. Scientific research has historically been one of the last domains where human expertise was considered irreplaceable. That assumption just got its first serious crack.

GPT-5.5 Launches Instead of GPT-6 And Still Takes the Lead

The model codenamed Spud shipped on April 23, not as GPT-6, but as GPT-5.5. OpenAI released it to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, with API access following on April 24. Greg Brockman called it “a new class of intelligence” and “a big step towards more agentic and intuitive computing.” Six weeks after GPT-5.4. GPT-6 remains unshipped. The wait that had the industry holding its breath resolved into something real and immediately consequential, just not the model the rumours named.

The capability profile is clearest in what GPT-5.5 does with a terminal. On Terminal-Bench 2.0, which tests complex command-line workflows requiring planning, iteration, and tool coordination, it scores 82.7 percent against Claude Opus 4.7’s 69.4 percent, a 13-point lead. That is the widest margin any model holds over its nearest competitor on any major agentic benchmark right now. On OSWorld-Verified, which measures whether a model can operate real computer environments autonomously, clicking, typing, and navigating interfaces, it reaches 78.7 percent, just ahead of Opus 4.7 at 78.0 percent. That is the first time OpenAI’s mainline model has matched Anthropic on full computer use, a category Opus has held outright since last autumn. The long-context retrieval jump is the other number worth keeping: MRCR v2 at 1 million tokens goes from 36.6 percent with GPT-5.4 to 74.0 percent with GPT-5.5. More than doubled. The model also uses 40 percent fewer tokens per Codex task than GPT-5.4, meaning tasks complete faster and cheaper on a per-outcome basis despite the API price doubling from $2.50 to $5 per million input tokens.

There is a catch worth naming. Claude Opus 4.7 still leads on SWE-bench Pro at 64.3 percent versus GPT-5.5’s 58.6 percent, and on MCP Atlas tool orchestration at 79.1 percent versus 75.3 percent. If your use case is resolving real GitHub issues or running complex multi-tool workflows, Anthropic is still ahead on those specific tasks. GPT-5.5 is the stronger model for agentic computer use and terminal work at scale. The race is not settled. It is now more clearly segmented by task than it was a week ago, which is more useful information than any single leaderboard number.

Why it matters

OpenAI spent six weeks building anticipation for GPT-6, shipped GPT-5.5 instead, and still took the agentic coding benchmark lead by 13 points. That says something about how far the baseline has moved. The model everyone was told to wait for still hasn’t arrived.

And that wraps up this week. Tune in next Monday, same time, for another deep-dive into the stories shaping the AI world.

The Sentinel lands in your inbox every Monday so you can catch up with the fast-moving AI space while sipping your morning coffee. Every detail that matters, none that doesn’t.

Here's what happened in the global AI market this week.

Stanford put hard numbers on how fast AI is outgrowing the tools to measure it, and it's frightening. Anthropic shipped its most powerful and capable model yet. NVIDIA just bridged AI and quantum computing. And the legal world got its first real warning about what AI conversations in court actually mean.

Here's everything you need to know before Monday gets the best of you.

Stanford's 2026 AI Index Is Out, And The Numbers Are Hard To Ignore

The most authoritative annual report on AI landed on April 13. The 2026 Stanford AI Index is 400 pages produced by an independent committee with no stake in any lab's outcome. And the picture it paints is one the industry has been quietly aware of but rarely said out loud: AI capabilities are racing forward, and the systems built to measure and govern them are getting left behind.

The capability numbers are striking. Coding benchmark performance on SWE-bench jumped from 60 to nearly 100 percent in a single year. AI models now outperform human baselines on PhD-level science and won gold at the International Mathematical Olympiad. Generative AI reached 53 percent of the global population faster than the personal computer or the internet ever did. And yet the same models that ace graduate-level physics can only read an analog clock correctly half the time. The "jagged frontier" is real, and the report documents it with more precision than anything else out there.

The geopolitical section is the one the industry will be quietly digesting for weeks. The US-China performance gap has essentially closed. Since early 2025 the two countries have been trading the top leaderboard spot back and forth, with the current lead sitting at just 2.7 percentage points. The US still dominates on investment, with $285.9 billion flowing into AI in 2025. But the number of AI researchers moving to the US has dropped 89 percent since 2017. And the most capable models today are among the least transparent, with major labs increasingly withholding the data that would let the outside world understand what they've actually built.

Why it matters

Stanford's Index is the closest thing to ground truth the field has that isn't produced by a lab with a financial interest in the result. This year's edition is the starkest documentation yet of the central tension in AI: capability is accelerating, and everything else is struggling to keep up.

Anthropic Releases Claude Opus 4.7, And It Has Raised The Bar By A Huge Margin

Anthropic released Claude Opus 4.7 on April 16, two months after Opus 4.6 and right on the rapid pace the company has kept across the 4.x generation. The headline improvements are in software engineering. A 13 percent gain on coding benchmarks, with users reporting the ability to hand off their hardest work, the kind that previously needed close supervision the whole way through, and actually trust the output.

But the more interesting changes are the ones developers will notice in production. Task budgets are new: the model now receives a token estimate for a full agentic loop and uses a running countdown to prioritize work and wrap up gracefully as the budget runs out. That changes how you design production systems. Instead of guessing when to cut a run short, you set a resource boundary and the model respects it. There's also a new high effort level and an /ultrareview command in Claude Code for when quality can't be compromised. Image resolution increased 3x, up to 3.75 megapixels. Pricing stayed the same at $5/$25 per million tokens.

The release also carries something quieter but significant. The cybersecurity safeguards in Opus 4.7 were directly informed by the Mythos Preview research, the same model that found thousands of zero-days and was deemed too dangerous to release publicly. Anthropic is threading the safety lessons from its most capable restricted model into the one you can actually use. Opus 4.7 is now live across Claude.ai, the API, Amazon Bedrock, and GitHub Copilot, where it replaces Opus 4.5 and 4.6 in the Copilot Pro+ model picker.

Why it matters

Task budgets alone change how reliable agentic systems get built. This isn't just a capability upgrade. It's a structural change in how AI agents behave in production, and that's the kind of update that compounds quietly across thousands of products built on Claude.

NVIDIA Launches Ising: The First AI Models Built Specifically For Quantum Computing

On April 14, NVIDIA announced Ising, the world's first open-source family of AI models designed specifically for quantum computing. The timing wasn't accidental. April 14th is World Quantum Day, and NVIDIA chose it to make a statement. The market heard it. Quantum computing stocks surged across the board, with IonQ up over 20 percent on the day.

The problem Ising is solving is one of the core reasons quantum computers have remained more theoretical than practical. Quantum processors need constant calibration to stay operational, tuning for hardware imperfections that shift over time, and real-time error correction, which means processing terabytes of qubit measurement data thousands of times every second. Both of these tasks have historically been slow, human-guided, and unable to scale. Ising's Calibration model automates the tuning using an agentic workflow, cutting setup time from days to hours. Ising Decoding handles the error correction, running 2.5x faster and 3x more accurate than existing approaches.

The models are open-source under Apache 2.0, pre-trained and ready to fine-tune, and already being used by Harvard, Fermilab, IQM Quantum Computers, Lawrence Berkeley National Laboratory, and the UK National Physical Laboratory. NVIDIA is positioning its GPUs as the classical computation layer that makes quantum systems actually run. The bridge between AI and quantum computing just got a lot shorter.

Why it matters

Quantum computing's real bottleneck has never been qubit counts. It's been calibration and error correction. By solving those two problems with open-source AI models, NVIDIA just removed the biggest practical obstacle between where quantum is today and where it needs to be to run useful applications.

Snap Cuts 1,000 Jobs And Explicitly Points At AI. The Stock Went Up 7 Percent.

On April 15, Snap CEO Evan Spiegel announced the company would cut 16 percent of its global workforce, around 1,000 full-time employees, and close at least 300 open roles. The reasoning wasn't dressed up in corporate language. AI advancements have made it possible for smaller teams to do the same work, and the company needs to reach profitability. The cuts are expected to reduce Snap's annualized cost base by more than $500 million by the second half of 2026.

The framing Spiegel used matters. He described AI as enabling teams to "reduce repetitive work, increase velocity, and better support our community, partners, and advertisers." He pointed to specific examples: small squads using AI tools to drive progress across Snapchat+, ad platform performance, and infrastructure efficiency. This is one of the most direct statements any major tech CEO has made, not a vague restructuring announcement, but a specific named mechanism for why people are being let go.

The market's reaction is the part worth sitting with. Snap shares rose 7 percent in pre-market trading when the news hit. The company joins a growing list including Meta, Oracle, and Amazon that have gone through significant cuts in 2026. The signal being sent, that AI-driven efficiency translates directly to shareholder value, is one that every technology company is now watching. The template is being set in real time.

Why it matters

This is one of the first times a major public company has cited AI this explicitly in a large-scale layoff and been rewarded by markets for doing so. That combination sets a template and an incentive that will be hard for other companies to ignore.

Spotlight

Krater Just Made Your AI Work While You Sleep

Most AI tools stop working the moment you close the tab. Krater's two new features change that.

Additions is a library of add-ons with three types: Personas, Prompts, and Apps. Apps are the interesting one. One click connects Gmail, Google Calendar, Notion, Slack, GitHub, Linear, or hundreds more. Once connected, Krater can read from and act on those services directly inside any chat. "Summarize my unread Gmail." "Find conflicts in my calendar next week." "Create a Linear ticket from this conversation." Connected apps stay linked to the account and are pinned to the top of the library so they're always one tap away.

Tasks is the scheduler. Any prompt worth repeating, a morning briefing, a weekly competitor scan, a daily inbox summary, can be turned into a recurring task with a name, a model, a cadence, and delivery preferences. Results land in a dedicated chat, as a notification, or straight to email. The real unlock is that Tasks run through the same engine as regular chats, which means they inherit every connected app. That means the task isn't limited to what the model knows. It reaches into the user's actual tools. Set "Every weekday at 8am, check my Gmail for anything urgent, cross-reference my calendar, and tell me what to focus on" and Krater does exactly that, every morning, before the user sits down. Apps give Krater hands. Tasks give it a clock. Put them together and the AI starts working for you while you sleep.

Try it

Additions and Tasks are live on Krater now. If your AI has been waiting for you to show up, this flips that around.

Section 230 Has Protected Platforms For 30 Years. AI Is Now Finally Breaking It.

The legal shield that has protected internet platforms from liability for three decades is starting to crack. This week, multiple cases advanced at once. The Massachusetts Supreme Judicial Court ruled on April 13 that Meta must face a state lawsuit over Instagram's design, finding that Section 230 does not protect how a platform is built and optimized, only what third parties post on it. A California jury had awarded $6 million in a similar case just weeks earlier. The pattern is becoming clear: courts are finding ways around the law that was supposed to make all of this impossible.

The AI angle is where things get particularly interesting. In the Northern District of California, courts are actively examining whether AI-generated advertising content changes the liability picture entirely. When a platform's AI systems shape, optimize, and effectively co-author how an advertisement is assembled and presented, courts are starting to ask whether the platform becomes the "maker" of that content rather than a neutral host. If that argument lands, Section 230 is irrelevant. Securities fraud law has no such immunity, and the exposure for Meta, Google, Snap, and anyone else running AI-optimized ad systems would be enormous.

Both Meta and Google have announced appeals. But the direction is clear. Multiple legal theories are bypassing Section 230 simultaneously. Product design liability, AI content generation, child safety. And they're landing now rather than a decade ago because AI has made the platforms' active role in shaping content impossible to ignore. The legal foundation of how the internet has operated for 30 years is under genuine pressure for the first time.

Why it matters

Section 230 is the bedrock of how internet platforms operate without facing unlimited liability. If AI-generated and AI-optimized content removes that protection, the entire business model of ad-supported platforms changes. And OpenAI, which just announced its own ad business, is building directly into this legal uncertainty.

Your AI Conversations With Claude Can Be Used Against You In Court. Lawyers Are Now Warning Clients.

A February ruling by a federal judge in New York has been making the rounds in legal circles this week, and the wave of warnings it triggered reached the mainstream. The case is United States v. Heppner. The defendant used Anthropic's Claude to prepare reports about his legal exposure and share them with his attorneys. His lawyers argued those conversations should be protected. The judge disagreed. Claude is not a lawyer. Public AI platforms have no confidentiality obligation. By accepting the terms of service, he had given up any reasonable expectation of privacy. The 31 AI-generated documents, including an outline of his own defense strategy, are now available to prosecutors.

The ruling created a split at the federal level. A Michigan judge reached the opposite conclusion the same day about a plaintiff's ChatGPT conversations, reasoning that AI tools are "tools, not persons" and that privilege requires disclosure to an adversary to be waived. The disagreement between federal courts means this question is unresolved nationally and heading toward higher courts. In the meantime, more than a dozen major law firms have issued advisories. New York-based Sher Tremonte now includes explicit language in client contracts warning that sharing a lawyer's advice with a chatbot could eliminate privilege entirely. The guidance ranges from choosing platforms carefully to writing specific language in prompts to document that queries are being made at a lawyer's direction.

The practical reality is broader than criminal defense. Anyone who has used a public AI platform to think through a legal situation, draft documents connected to a dispute, or discuss anything that could end up in litigation needs to understand that those conversations may be discoverable. The case also clarifies a distinction that enterprise AI has been quietly building toward: closed, corporate systems with real confidentiality protections sit in a different legal position than public chatbots. That difference just became a lot more concrete.

Why it matters

This is the first ruling of its kind in the US, and it's now generating national legal guidance. The line between a helpful AI tool and a liability risk just got clearer, and it runs directly through the terms of service most people click through without reading.

GPT-6 Still Hasn't Launched. And The Silence Is Becoming A Bit Too Loud

Pre-training for GPT-6, internally codenamed "Spud", was completed on March 24 at OpenAI's Stargate data center in Abilene, Texas. Sam Altman confirmed this on X and described the launch as "a few weeks" away. Greg Brockman called it "two years of research" and "not an incremental improvement." A leaked rumor named April 14 as the date. That date passed without a word. No announcement, no blog post, no Altman tweet. As of Friday, the model has not shipped.

The context makes the silence notable. Every major competitor shipped a flagship model in April. Anthropic released Opus 4.7. Google's Gemini 3.1 Ultra is in public hands. Meta's Muse Spark has been out for two weeks. OpenAI is the last major lab still holding back, and the longer it takes, the more of the "frontier model" narrative gets eaten by everyone else. Prediction markets have trimmed the probability of a launch by April 30 from 78 percent to around 43 percent. The most credible explanation is safety evaluation, a model that Brockman described as "not an incremental improvement" requires more extensive red-teaming, and that takes time.

The leaked specification remains unverified. 40 percent performance improvement over GPT-5.4, a 2 million token context window, native multimodality. OpenAI has published no official model card, no pricing, nothing. What's confirmed is that the model exists, pre-training is complete, and the company is facing more competitive pressure to ship than at any previous release. A May or early June window is now the most reasonable estimate. The market is waiting, and the wait itself is now part of the story.

Why it matters

GPT-6 will be the most scrutinised AI release in history. Every week it doesn't ship while competitors have already moved, the expectations climb higher. How it lands relative to those expectations will define OpenAI's competitive position for the next model generation.

Anthropic's Claude Cowork Is Now Enterprise-Ready, And That Closes An Important Gap

Claude Cowork became generally available on April 16, launching on macOS and Windows via the Claude Desktop app alongside the Opus 4.7 release. The timing was deliberate. A new capability and the infrastructure to deploy it at scale shipping on the same day is not a coincidence. Cowork now includes role-based access controls, group spend limits, detailed usage analytics, Claude Cowork in the Analytics API, and OpenTelemetry support for enterprise observability tooling.

The gap this is designed to close is one that rarely makes headlines but explains a lot about why enterprise AI adoption has been slower than the capability numbers would suggest. It's almost never the model. It's the governance layer. Enterprise procurement processes require the ability to see who is using what, set and enforce spending boundaries, manage access by role, and pull usage data into existing reporting infrastructure. Models that can't satisfy those requirements don't get purchased at scale, regardless of how impressive they are in a demo. Cowork's governance features are the answer to the question that procurement teams have been asking for the past two years.

The Analytics API integration is worth noting separately. Enterprise clients can now pull detailed engagement and adoption data programmatically, feeding AI usage into the same dashboards they use for every other enterprise tool. Combined with the Opus 4.7 upgrade, this week's releases position Anthropic to close deals that have been sitting in procurement for months. The distance between what Claude can do and what organisations are actually using it for has been narrowing. This week it narrowed significantly.

Why it matters

Enterprise AI adoption is bottlenecked on governance, not capability. Cowork GA removes the blockers that have been holding back large-scale deployments. The companies that move on this quickly will compound their AI advantage over the ones still waiting for procurement sign-off.

And that wraps up this week. Tune in next Monday, same time, for another deep-dive into the stories shaping the AI world.

The Sentinel lands in your inbox every Monday so you can catch up with the fast-moving AI space while sipping your morning coffee. Every detail that matters, none that doesn't.

Here's what happened in the global AI market this week.

AI found open cracks, and that raised some serious security concerns. OpenAI had a $122B funding round, but it goes beyond just numbers. Benchmarks are no longer enough. AI is getting ahead of human-made performance tests. And many more intriguing stories.

Here's everything you need to know before Monday gets the best of you.

Claude Found Thousands of Zero-Day Vulnerabilities. Anthropic Kept The Matter Among Trustees

This wasn't a demo or a controlled experiment. Anthropic's Claude Mythos Preview autonomously discovered thousands of previously unknown security vulnerabilities a.k.a zero-days across major operating systems and browsers. The scope was significant enough that Anthropic didn't publish the findings. Instead, it launched Project Glasswing: a coordinated disclosure effort with major tech companies to quietly fix the flaws before anyone else could find and exploit them.

The significance of what happened here is easy to underestimate. Finding zero-days has historically required elite human researchers going through code day in and day out for weeks and sometimes, even months. Claude did it at scale, autonomously, in a fraction of the time. That's not just a capability improvement. It just changed what was thought possible in cybersecurity. And now that everyone knows the capability to find these flaws at such a huge scale exists, the danger of its weaponisation has gone sky-high.

Anthropic made the right call by coordinating disclosure and not sweeping things under the rug. But the conclusion of this scenario is hard to ignore. Once this technology exists at one place, it's a matter of time before it is replicated. The security industry's entire threat model just changed.

Why it matters

Anthropic won this round. But the cat is now very much aware of the bag. Every future AI researcher now knows this is possible.

OpenAI's $122B Round: The Numbers Are Real, But They're Not Everything

OpenAI closed a funding round that valued it at $852 billion, a number that would have seemed like a fairytale two years ago. But the structure of the deal is worth understanding before being blown away by mere numbers. Of the $122B raised, Amazon, Nvidia, and SoftBank account for $110B. Much of that isn't clean equity. It's vendor deals and contingent capital, meaning a considerable portion is tied to commercial arrangements rather than pure investment.

Then there's the side deal: OpenAI is reportedly negotiating a ~$10B private equity joint venture that comes with a guaranteed minimum return of 17.5%. That's an unusual structure for the most prominent AI company in the world. Guaranteed returns suggest that the equity story was insufficient to attract capital on standard terms and the investors needed more certainty.

Separately, OpenAI confirmed it expects its advertising business to generate $2.5 billion in revenue this year, with an estimate of $100 billion by 2030. The company is no longer just a model provider. It's building a media and ad business and looking to compete directly with big names like Google and Meta, both of whom it currently relies on for infrastructure and distribution.

Why it matters

The world's most valuable AI company is paying guaranteed returns to raise capital and building an ad business to sustain itself. That's not a startup dynamic anymore. It's the behavior of a scaled platform company that is in desperate need of revenue.

Our Tests Are No Longer Good Enough to Keep Up With AI

Time Horizon suite, one of the most rigorous evaluation frameworks in existence, is now at the ceiling. Frontier models are saturating it, meaning they're performing at or near the ceiling on almost all of the tasks, making it impossible to distinguish between different capability levels at the top end.

The problem isn't just that one benchmark is getting old. It's that building replacements is genuinely hard. New benchmarks require expensive human verification, careful task design, and time. A lot of time. And AI capabilities are outpacing it way too fast for the tests to keep up. Researchers now warn that by mid-2027, no benchmark created in 2026 or earlier will reliably be able to rule out dangerous capabilities in frontier systems. You can't govern what you can't measure, and right now, measurement is falling behind.

This matters beyond AI safety circles. Regulators, enterprise buyers, and governments rely on benchmark scores to make decisions about which AI systems are safe to deploy. If those scores can no longer differentiate between systems that are safe and systems that aren't, the entire evaluation infrastructure that decides the fate of the models goes extinct.

Why it matters

If we can't measure what AI can do, we can't control it. The infrastructure we use to govern AI is quietly and quickly becoming outdated and nobody has a clean solution yet.

Drug Discovery, $10B In Revenue And More In Anthropic's Historic Week

While Project Glasswing took the headlines, Anthropic had a week that would have been front page in the news in either way. First, the company acquired Coefficient Bio, a 10-person startup founded by ex-Genentech researchers, for $400M in stock. Coefficient Bio was using AI to accelerate drug discovery and biological research, and the team will now integrate into Anthropic's health and life sciences division. Drug discovery is now on Anthropic's agenda.

At the same time, the company signed a deal with Google and Broadcom for multiple gigawatts of next-gen TPU capacity starting 2027. This goes beyond renting cloud compute. Anthropic is securing the physical hardware foundation needed to train and serve future generations of models at a scale that matches its ambitious plans.

And quietly, Anthropic hit $10 billion in annual revenue, beating Shopify, ServiceNow, or almost any software company in history to this milestone. It hasn't been four years since founding. That means despite fierce competition, Anthropic has been outpacing its competitors on all major fronts, quickly becoming a major player in the AI market and now has the revenue numbers to show for it.

Why it matters

Anthropic is no longer just a second option to ChatGPT. It's becoming a full-stack technology company with healthcare branches, independent compute, and revenue that has to be taken seriously.

Reputed Researchers Are Now Cutting Their AI Timeline Estimates By Half

Researcher Eli Lifland updated his probability estimates this week, placing nearly 2x higher odds on full AI R&D automation by the end of 2028.

The sudden shift isn't the result of any major breakthroughs in creativity. It's driven by something boring but powerful. AI can now perform large-scale, specification-driven software engineering faster and cheaper than human engineers. The tasks that require straight-forward execution. Refactoring, testing, debugging, implementing specified features. The kind of work that takes up a major chunk of the software team's time.

What makes this consequential is that a large portion of AI R&D itself consists of exactly this type of task. If AI can do that work faster and cheaper, it can multiply the rate of its own development. The loop begins. AI improves AI. And that loop starting means that humans can no longer predict how advanced AI will be 2 or 3 years into the future. Lifland's update isn't a fire-hazard alarm. But it's something worth keeping an eye on.

Why it matters

The "AI speeds up AI" dynamic is no longer theoretical. When reputed forecasters update their timelines by 2x, it's worth paying attention to what changed in their reasoning, not just the number.

Meta's Muse Spark Is the Most Competitive Model They've Ever Released

Meta launched Muse Spark, multimodal reasoning with visual chain-of-thought, tool use, multi-agent orchestration, and the early benchmarks are quite impressive. Impressive enough to put it in competition with the leading frontier models across multiple domains, not just the categories where Meta has always been big. This is positioned as part of the company's push toward what Zuckerberg has been calling "personal superintelligence."

The open-source question is still there. Meta confirmed new models are coming and some will eventually carry an open license. Whether that's Muse Spark or other variants in the same family hasn't been confirmed. Zuckerberg has been vocal about open-source AI as a long-term differentiator, and Llama's track record has been consequential in shaping the open model ecosystem. Whether Muse Spark itself gets open-sourced, or whether that designation applies to other future models in the lineup, hasn't been confirmed.

What's changed with this release is the capability profile. Previous Meta models were strong in specific areas. Muse Spark appears to be competitive across much more. Reasoning, vision, tool use, and orchestration, which puts it in a different ballpark than what Meta was in six months ago.

Why it matters

Meta has historically been the company most willing to open-source capable models. If Muse Spark-tier performance eventually ships under an open license, the accessible capability baseline for anyone building on open models moves up significantly.

Perplexity Has Grown 50%In Revenue And Also Gained Access To Your Bank Account

Perplexity reported 50% revenue growth after its strategic shift from AI-powered search to AI agents. The company's bet is paying off as more and more users now want AI that does things, finds things. The search option still exists, but the company's growth is now being driven by agent capabilities.

Alongside the revenue boost, Perplexity launched a personal finance dashboard powered by Plaid integration. Users can now link checking accounts, savings accounts, credit cards, and loans directly to Perplexity and receive AI-driven analysis across all of it. Spending breakdowns, liability tracking, net worth calculation and much more. Traditional personal finance apps show you data. Perplexity asks what you want to understand about it and tells you.

Perplexity is building a product that covers search, agents, and now personal finance all under the same interface. The underlying bet is that users will anchor their daily AI usage around Perplexity the way they anchored their daily searches around Google. The 50% revenue growth suggests people want the product enough to pay for it. The product experience is good enough that users are connecting sensitive financial data to it. That level of trust is not easy to earn and even harder to maintain at scale. But with the way Perplexity is moving, the next 12 months are going to be interesting.

Why it matters

Perplexity is quietly building an AI-native super-app. The personal finance integration shows how far beyond search the company is willing to go to capture habitual daily usage.

Anthropic's Updates Just Completely Changed How You Build With Claude

Claude just made 3 new updates each targeting a very real problem.

Managed Agents separates the agent interface from the model layer, meaning systems built on Claude can evolve with new model versions without requiring architectural rebuilds. For anyone who has managed a production AI system through a major model update, the value of that stability is enormous. It removes a category of technical debt that has made long-term AI infrastructure planning very difficult.

Claude Cowork has been updated to enterprise readiness, adding role-based access controls, group spend limits, detailed usage analytics, and Zoom integration. Companies like Zapier and Airtree are already using it for project management and operational workflows. The enterprise layer matters because it's where Anthropic's $10B revenue is actually being generated, and this update just made that level even stronger.

The Advisor API Tool is the most technically interesting of the three. It allows developers to configure Opus as a reasoning advisor while running Sonnet or Haiku as the executor for the actual work. You get Opus-level thinking on the important decisions, at Sonnet-level pricing on the execution. The fundamental tradeoff that has forced product teams to choose between quality and sustainability now has a structural solution. This will change product decisions at hundreds of companies.

Why it matters

The Advisor Tool alone could meaningfully reduce the per-task cost for AI product builders without sacrificing output quality. That's the kind of infrastructure update that compounds quietly across thousands of products.

And that wraps up this week. Tune in next Monday, same time for another deep-dive into the major AI news.

The Sentinel lands in your inbox every Monday so you can catch up with the fast-moving AI space while sipping your morning coffee. Every detail that matters, none that doesn't.

The artificial intelligence landscape in week 15 of 2026 continues to accelerate with remarkable intensity, featuring groundbreaking model releases from industry leaders, significant security concerns that demand immediate attention, substantial corporate funding announcements, and evolving enterprise deployment strategies. This week marks a pivotal moment where the theoretical capabilities of advanced AI systems are transitioning into concrete business applications, while simultaneously raising critical questions about safety, security, and responsible deployment that cannot be ignored by organizations seeking to harness these powerful technologies.

Major Model Releases and Technical Breakthroughs Reshape the Competitive Landscape

The week opened with Meta’s much-anticipated announcement of Muse Spark, the first production model from its newly established Meta Superintelligence Labs. This multimodal reasoning model represents a significant recovery for Meta following the poorly received Llama 4 release in April 2025, and it demonstrates the company’s ability to rebuild its AI capabilities from the ground up. According to Meta’s technical documentation, the team accomplished this feat by completely reconstructing the company’s AI stack over nine months, achieving comparable capabilities to its previous iterations while using “over an order of magnitude less compute” than Llama 4 Maverick.

Muse Spark’s performance metrics are genuinely competitive with leading frontier models across numerous domains. The model achieved a score of 42.8 percent on HealthBench Hard, surpassing both Anthropic’s Opus 4.6 and Google’s Gemini 3.1 Pro, while remaining competitive with OpenAI’s GPT-5.4. The model incorporates native multimodal capabilities, enabling it to process both text and images seamlessly, and it supports what Meta describes as a “contemplating” or “thinking” mode that deploys subagents to reason about different aspects of tasks in parallel. However, Meta acknowledged performance gaps in long-horizon agentic systems and coding workflows, areas where it intends to focus continued investment. The company has rolled out Muse Spark across its entire product ecosystem, including the Meta AI app, meta.ai, WhatsApp, Instagram, Facebook, Messenger, and Ray-Ban AI glasses, with private preview access available to select partners through an API.

Beyond Meta, the week saw substantial activity in the open-source model space. Google released Gemma 4, its most capable open-weight model family designed specifically for advanced reasoning and agentic workflows. The Gemma 4 series delivers what Google describes as “unprecedented intelligence-per-parameter,” with multiple variants spanning from edge devices to data centers. These models incorporate advanced reasoning capabilities and native support for agentic workflows, representing Google’s commitment to competing effectively in the open-source AI race.

Anthropic also advanced its Claude product line with significant updates throughout Q1 2026, shipping over 35 new features in roughly 90 days. The company’s Claude Opus 4.6 model, released in February, continues to demonstrate superior performance on complex reasoning tasks and long-context processing challenges. In March, Anthropic expanded access to Claude’s 1-million-token context window at standard pricing and doubled off-peak usage limits, making sophisticated long-context capabilities more accessible to enterprise customers.

Mistral AI unveiled Mistral Small 4, described as the first unified model combining the capabilities of its previous flagship offerings across reasoning, multimodality, and coding. With 119 billion total parameters and 6 billion active parameters per token, the model supports a 256,000-token context window and delivers what Mistral characterizes as best-in-class efficiency gains, achieving a 40 percent reduction in end-to-end completion time in latency-optimized configurations.

Critical Security Vulnerabilities and the Rise of AI-Driven Cybersecurity Threats

This week brought sobering developments regarding AI capabilities in cybersecurity that warrant serious organizational attention. Anthropic released details about Mythos, its new AI model specifically designed for identifying and exploiting software vulnerabilities. The company states that Mythos demonstrated such advanced capabilities in autonomously discovering and testing vulnerabilities that it decided to restrict access to a carefully selected group of major technology companies whose foundational software supports many other digital services. According to Anthropic’s announcement, Mythos successfully reproduced and exploited vulnerabilities in over 80 percent of test cases, representing a capability threshold that prompted the company to adopt a phased rollout approach designed to give defenders a head start before similar capabilities become widely available.

Security researchers have emphasized that the significance of Mythos lies not merely in its ability to identify vulnerabilities within code segments researchers pre-identified, but rather in its capacity to autonomously locate vulnerabilities buried within millions of lines of code and subsequently verify that these vulnerabilities constitute genuine exploitable weaknesses. This autonomous capability makes Mythos fundamentally different from and more dangerous than smaller, openly available models that can identify vulnerabilities only when researchers know where to direct the model’s attention.

Smaller models may eventually achieve comparable results to Mythos, according to cybersecurity experts, but they require substantially greater technical skill, careful prompt engineering, and purpose-built tooling to accomplish similar feats. Mythos, by contrast, dramatically lowers the barrier to entry for conducting sophisticated, large-scale cyberattacks, potentially placing such capabilities within reach of malicious actors lacking deep technical expertise. OpenAI is also reportedly preparing a cybersecurity-focused model internally known as “Spud” with capabilities potentially matching Mythos, planning a similarly phased rollout strategy to give defenders access before broader deployment.

In response to these emerging threats, Anthropic established Project Glasswing as a coordinated effort to provide advanced cybersecurity capabilities to defenders while managing risks. The company is offering up to $100 million in usage credits to enable select infrastructure organizations to strengthen their defenses and is actively coordinating with government stakeholders to ensure appropriate oversight. Despite these mitigation efforts, security experts caution that much of what Mythos can accomplish may already be possible with smaller, openly available models for those with sufficient expertise, though automation of the vulnerability discovery and exploitation pipeline represents a significant and concerning escalation.

Financing Frenzy: Record AI Investment Drives Unprecedented Consolidation

The venture capital landscape in 2026 has become almost unrecognizable compared to previous years, with Q1 establishing records that dwarf any prior funding period in history. Global venture investment reached $300 billion across 6,000 startups in the first quarter of 2026, representing an increase of over 150 percent quarter-over-quarter and year-over-year. This quarterly sum alone accounts for nearly 70 percent of all venture capital deployed throughout 2025, and surpasses the entire full-year investment totals for every year before 2018.

The capital concentration in AI has been extraordinary, with $242 billion—representing 80 percent of total global venture funding in Q1—directed to artificial intelligence companies and applications. This represents a dramatic increase from Q1 2025, when AI accounted for 55 percent of global venture funding. U.S.-based companies captured $250 billion of this total, representing 83 percent of global venture capital, up substantially from 71 percent in Q1 2025.

Four of the five largest venture funding rounds ever recorded in history were closed during this quarter, with frontier AI laboratories and autonomous vehicle developers commanding the vast majority of capital. OpenAI secured $122 billion in funding, achieving a valuation of $500 billion, establishing it as the most valuable artificial intelligence company. Anthropic closed a $30 billion funding round valuing the company at $183 billion, with the company’s run-rate revenue having surpassed $30 billion as of early April, more than triple the approximately $9 billion reported at the end of 2025. Remarkably, the number of business customers spending more than $1 million annually on an annualized basis has doubled from over 500 in February to exceeding 1,000 in April.

xAI raised $20 billion to support its advanced AI development initiatives, while Waymo secured $16 billion in funding. These four investments alone accounted for $188 billion, representing 65 percent of all global venture investment in Q1. This concentration of capital reflects investor conviction that frontier AI capabilities represent the most significant commercial opportunity of our time, though it also raises questions about the sustainability and distribution of AI benefits across the broader economy.

Anthropic Expands Compute Infrastructure to Support Explosive Customer Demand

Anthropic announced a groundbreaking partnership with Google and Broadcom to secure multiple gigawatts of next-generation TPU computing capacity expected to come online beginning in 2027. This represents the company’s most significant compute infrastructure commitment to date and reflects the extraordinary demand being experienced from Claude customers worldwide. The vast majority of this new computing capacity will be sited within the United States, making this partnership a major expansion of Anthropic’s November 2025 commitment to invest $50 billion in strengthening American computing infrastructure.

This expansion deepens Anthropic’s existing collaboration with Google Cloud while also strengthening the company’s relationship with Broadcom. Anthropic has emphasized its deliberate strategy of training and deploying Claude across diverse hardware platforms, including AWS Trainium accelerators, Google TPUs, and NVIDIA GPUs, enabling the company to match specific workloads to the chips best suited for their execution. This hardware diversity translates to superior performance and greater resilience for customers who depend on Claude for mission-critical work, while Claude remains the only frontier AI model available to customers across all three of the world’s largest cloud platforms—Amazon Web Services through Bedrock, Google Cloud via Vertex AI, and Microsoft Azure through Foundry.

OpenAI’s Pricing and Product Strategy Signals Intensifying Competition

OpenAI announced a new $100-per-month Pro plan for ChatGPT, directly challenging Anthropic’s existing $100-monthly pricing tier for Claude. This new pricing tier is expressly designed to provide developers with substantially greater coding capacity through OpenAI’s Codex tool, especially during high-intensity work sessions where usage limits become constraining. OpenAI emphasizes that Codex delivers more coding capacity per dollar across paid tiers compared to Claude Code, with the difference being most pronounced during active coding use. According to OpenAI, more than 3 million people globally are using Codex every week, up five-fold in the past three months, with usage growing more than 70 percent month-over-month.

The company is offering elevated Codex limits on the $100 plan through May 31, 2026, though users should anticipate that such promotional limits will normalize after that date. The $200-per-month plan remains available for the most demanding workflows, offering 20-times higher limits than the Plus tier and described as sufficient to support “your most demanding workflows continuously, even across parallel projects”. OpenAI’s aggressive pricing and feature positioning reflects the intense competitive dynamics now characterizing the AI assistants market, with companies racing to capture developer mindshare and establish long-term customer relationships.

Enterprise Adoption Accelerates Despite Governance Challenges

Enterprise AI adoption continues to accelerate across industries, with worker access to AI tools increasing by 50 percent in 2025, and expectations for continued expansion remain high. Organizations report that the number of companies with 40 percent or more of projects in production is expected to double within the next six months. However, this rapid deployment is occurring amid concerning governance gaps, with only one in five companies maintaining mature governance models for autonomous AI agents.

Gartner projects that 40 percent of enterprise applications will embed task-specific AI agents by the end of 2026, up from fewer than 5 percent one year ago. This explosive projected growth reflects fundamental organizational recognition that AI agents have transitioned from experimental tools to essential business infrastructure. Enterprise leaders identify improving productivity and efficiency as the primary achieved benefits from AI adoption thus far, with two-thirds of organizations reporting such gains.

However, enterprises simultaneously report feeling less prepared operationally for AI adoption when compared to previous years, despite more companies believing their strategies are highly prepared. Specific areas of concern include infrastructure readiness, data quality and governance, risk management, and talent availability. This preparedness gap suggests that while organizations are moving quickly to deploy AI agents across business functions, many may be doing so without adequate structural preparation and governance frameworks.

Regulatory Framework Developments and Policy Discussions

The White House released a comprehensive national AI legislative framework in March 2026 proposing federal preemption of fragmented state AI laws. The framework recommends that Congress establish federal standards while preempting state laws that impose what the administration characterizes as “undue burdens,” while preserving states’ traditional police powers to enforce generally applicable laws protecting children, preventing fraud, and safeguarding consumers. The framework further recommends that states be prohibited from regulating AI development, which the administration describes as “an inherently interstate phenomenon with key foreign policy and national security implications”.

On intellectual property matters, the White House framework supports allowing courts to resolve whether AI training on copyrighted material constitutes fair use without Congressional intervention, while recommending that Congress consider enabling collective licensing frameworks that allow rights holders to negotiate compensation from AI providers. Additionally, the framework calls for commercially reasonable age-assurance requirements for AI platforms likely to be accessed by minors, affirming that existing child privacy protections apply to AI systems, and recommends establishment of federal rights protecting individuals from unauthorized AI-generated digital replicas of their voices and likenesses.

Senators Adam Schiff (D-California) and John Curtis (R-Utah) introduced the Copyright Labeling and Ethical AI Reporting (CLEAR) Act, which would establish mandatory notice requirements for companies developing AI models trained using copyrighted works. The legislation would direct AI developers to submit notice to the U.S. Copyright Office of all copyrighted works included in training datasets, with such notice required to be filed 30 days before commercial release of the AI platform. Generative AI developers failing to comply could face civil penalties of $5,000 per instance of failed notice, with maximum penalties of $2.5 million per case without limitation to other copyright remedies.

Workforce Disruption and the Reality of AI-Driven Layoffs

The technology sector experienced significant workforce reduction in Q1 2026, with 78,557 workers laid off from January through April 2026, with more than 76 percent of affected positions concentrated in the United States. Approximately 37,638 of these cuts—representing 47.9 percent of total tech layoffs—have been attributed to reduced demand for human workers due to AI and workflow automation capabilities. Major contributors to these layoffs included Meta’s 1,500-person reduction from its Reality Labs division, Block’s 40-percent workforce reduction affecting over 4,000 employees, and numerous smaller but significant reductions across the industry.

Cognizant’s Chief AI Officer Babak Hodjat cautioned that the full impact of modern AI technologies on the workforce will not become apparent for more than a year, and suggested that some companies are using AI as a “scapegoat” to justify workforce reductions they would have undertaken regardless of technological developments. However, Hodjat acknowledged that real AI-driven layoffs will likely occur, with companies potentially seeing substantial productivity gains from AI within six to twelve months of deployment. This candid assessment suggests that while some current layoffs may be conflated with AI displacement, meaningful productivity-driven workforce adjustments are reasonably anticipated as AI systems mature and organizational practices optimize around their capabilities.

Looking Ahead: The Trajectory for Enterprise and Consumer AI

The week’s developments collectively demonstrate that 2026 is consolidating the transition from AI experimentation to operational deployment across enterprises, with frontier model capabilities becoming increasingly sophisticated while simultaneously generating novel security and governance challenges. The record capital flows into AI companies reflect investor conviction that these systems represent genuinely transformative business opportunities, though the sustainability of current valuation multiples and funding levels remains uncertain. Organizations seeking to deploy AI responsibly must simultaneously accelerate adoption while establishing mature governance frameworks, security practices, and workforce transition strategies to navigate this period of unprecedented technological change.

The convergence of increasingly powerful models, aggressive enterprise deployment, emerging security threats, and fragmented regulatory frameworks creates both extraordinary opportunities and substantial risks for organizations navigating the 2026 AI landscape. Success will depend on balancing competitive urgency against the essential work of building trustworthy, secure, and ethically sound AI systems that deliver genuine business value while managing the legitimate societal concerns this technology raises.