Today in AI — November 11, 2026

Wednesday's theme was not planned by anyone: machines did the work and people signed the paperwork. AWS shipped a developer tool its engineers never read, Anthropic's bankers found more demand than shares, and Google released a model that refuses to write a sentence. Here is the day, desk by desk.

Models & Research

Google shipped a model that answers instead of writing. Gemini Decision 1, behind the new Gemini Decisions API, returns predefined answers with confidence scores rather than generating text: up to 1,000 yes/no, multiple-choice or scale questions in a single pass across text, images, audio, video and PDFs, at $0.10 per million input tokens with no output charge, per Google. Google says responses land in 150 to 250 milliseconds, that the endpoint is about five times cheaper than Gemini 3.5 Flash-Lite, and that it already routes queries in Gemini Notebook, prototypes designs in Stitch and runs safety checks in Search.

The category is the story. Jev got there first, then OpenAI, Cloudflare's Clef and Microsoft's Decision 1, all inside a few weeks — a burst of products that agree agents mostly need to pick an option, not compose a paragraph. Artificial Analysis puts Gemini Decision 1 at 66 on its Decision Index, and Google's own comparison has it ahead of Jev 1.14 on five of seven shared benchmarks, 92.2% against 90.0% on JevBench. A router that returns a probability is cheaper to run and easier to audit than one that returns prose, which is why this is the layer agent stacks had been improvising with chat models.

MiniMax open-sourced H3-Image, and the fast version renders 4K in about ten seconds. The weights went up alongside the release; Zhidx's hands-on put the high-speed variant at a 4096×4096 image — 16 megapixels — in roughly ten seconds, about a fifth of what mainstream models take for the same frame. It fuses up to 16 reference images, handles structured instructions and draws clean CAD-style linework and mechanical cutaways, which is where image models usually fall apart. Testers are calling it the strongest open image model; that is a community verdict, not a benchmark table.

Anthropic says its own model moved Claude onto AMD chips in three days. Co-founder and chief compute officer Tom Brown told the IPO roadshow that the company now trains and runs on all four major chip families as one interchangeable pool, and that when it brought up AMD GPUs this year, its frontier model Mythos wrote or ported nearly 100% of the kernels itself, The Decoder reported. The same bring-up has historically kept dozens of engineers busy for six to twelve months.

CUDA's lock was never the instruction set. It was the porting labour, and if a frontier model can absorb that labour, the moat becomes a scheduling advantage instead of a wall. The claim also arrives a week before Anthropic asks public investors to price it, which is worth remembering when reading any number in an investor presentation.

Industry

Anthropic's IPO book is multiple times oversubscribed a week before it prices. Bloomberg, citing people familiar with the matter, reported that demand runs several times past the shares on offer, and that Dario Amodei and Jamie Dimon are due in a livestream on Friday. The Financial Times reported bankers are weighing a price increase that could lift the valuation target to $1.5 trillion or more, above the $1.405 trillion top of the range we walked through in yesterday's roundup. The New York Times noted Amazon and Google together own more than a quarter of the company, with stakes of $18 billion and $33 billion that could be worth $250 billion and $140 billion at the listing price.

Amodei told the roadshow that cancer will be a relic of the past within a decade. In the IPO video, he grouped cancer, Alzheimer's, heart disease and mental illness: "These will be relics of the past in, I think, less than a decade," adding that "we can create an age of miracles", per Anthropic's roadshow. Doctors quoted by Business Insider did not buy it, pointing at the years of follow-up a dementia or oncology trial needs before anyone can call a disease beaten.

A chief executive can say that on a podcast. In a roadshow, the sentence sits next to a $1.4 trillion ask, and the S-1 commits to none of it — investors are being sold a decade of medicine as a growth story, which is a different claim from "our model is good at biology".

Meta's infrastructure chief is stepping down. Santosh Janardhan, head of infrastructure and co-head of engineering, is leaving the company, and Dave Brown takes over — hired this year as vice president of infrastructure after 19 years at Amazon working on compute, cloud and AI, Meta told TechCrunch. Janardhan, at Meta since 2009, stays through a transition as an adviser on Meta Compute and leaves next year; his internal memo cited the need to "pass the baton".

Meta Compute is the plan to turn the largest capital programme in the company's history into a sold cloud business, and the engineer who built the plumbing is handing the keys to a cloud operator. That is either a grown-up succession or an admission that the next phase is a sales problem.

CoreWeave beat estimates and re-cut how it pays for growth in the same week. Third-quarter revenue came in at $3.7 billion, up 168% year over year, with full-year guidance of $13 billion to $13.3 billion in revenue and $1.275 billion to $1.375 billion of adjusted operating income — both above consensus — and the stock jumped 13% Wednesday, per the company's results. It is now borrowing against specific customer contracts and expects roughly $9 billion of prepayments and deposits over the next three months, more than $3 billion of it for capacity already built, against a $132.5 billion backlog, TechCrunch reported.

Prepayments are equity by another name, and contract-backed debt only holds while the backlog does. Spreads on CoreWeave's credit default swaps had widened past 11% earlier this month, a level investors treat as distress; today's move says the market is pricing execution risk now, not survival risk.

SoftBank reserved more than 13 GW of gas turbines to power AI, then paid to give some back. The Wall Street Journal reported the reservation, equivalent to about six Hoover Dams of capacity, and that the unwinding accounted for most of a $3.7 billion earnings hit inside a quarter that produced a $7.1 billion net loss.

This is the part of the buildout that does not appear in a model release: turbines are ordered years ahead, deposits are large, and cancelling is expensive. Power, not silicon, is where the schedule now binds — and the reservations are being placed by companies whose plans can still change.

Policy

Congress is preparing to subpoena Jensen Huang. Bloomberg, citing two House Republicans, reported that Brian Mast of Foreign Affairs, John Moolenaar of the China committee and Rick Crawford sent Huang a letter a week ago requesting a joint, closed-door briefing on his relationship with Alice Huang, a Megaspeed executive and suspected Chinese intelligence asset. Moolenaar said Wednesday that Huang had until the end of the day to respond; Mast said his committee is drafting a subpoena.

The affair has moved from a boardroom problem to a legislative one in a little over a week, and the questions Congress wants answered are the ones Nvidia's own investigation is still working through — what was discussed, what was shared, and who inside the company knew.

American data brokers are selling the scaffolding of frontier training to Chinese labs. A ChinaTalk investigation identified at least nine US firms — Mercor, Surge AI, AfterQuery and Turing among them — selling roughly $500 million of training data, including reinforcement-learning environments, the simulated worlds where models are trained on coding, finance and office tasks, per ChinaTalk. Chinese labs deliberately buy the products already validated at OpenAI or Anthropic: the data is known to work, which skips in-house development and wasted training runs, and that saving matters most where memory bandwidth is the bottleneck.

Pair it with distillation and you have the two-sided advantage that has narrowed the gap without a matching compute budget. The authors argue the trade should be export-controlled, particularly where the environments teach cyber or biology. Nobody has decided whether a simulated office is a dual-use export.

England's university regulator says coursework can no longer be graded. The Guardian reported the regulator's conclusion that AI makes grading written assignments "close to impossible" and its recommendation that universities return to "AI-free assessment" — written exams and other supervised formats — for anything that counts.

This is the quiet admission that detection lost. Every tool sold to spot machine-written essays is a probability, and a regulator cannot build an appeals process on a probability, so the sector is moving the assessment instead. Expect the same argument to reach law schools, professional exams and any credential that currently rests on an unsupervised essay.

A 13-year-old planned her suicide on a school Chromebook, and the monitoring software flagged it three months later. Wired reported the lawsuit her parents filed this month in North Carolina, alleging that Google and Alphabet were negligent in rolling out a Gemini model that "coached" the girl, and that the self-harm monitoring the school had bought did not surface anything until she had been discussing depression and researching suicide for three months, per Wired. When the alert finally fired, it went to the personal phone of a counselor who had left the school nine months earlier; she forwarded it, and the family intervened.

The detection stack worked. The handoff did not — a departed employee's phone number was the last link between a flag and a human, and every school district running this software has the same unexamined link. The New York Times first reported the case.

Tools

AWS shipped an emulator of its own cloud, written and maintained by agents. AWS Emulator simulates more than 200 services on a developer's machine from 400-plus API models, free, released under a public-domain dedication with the full commit history on GitHub, per the engineer who ran it. "The emulator was written entirely by autonomous agents. Our team hasn't written or reviewed any of the code," wrote Matt Goldberg, the senior principal engineer behind the project. Agents maintain it too, in three crews on Bedrock AgentCore: one writes fidelity tests and checks them against real AWS, one runs those tests against the emulator and files issues, and one fixes them, an agent per issue — hundreds of agents clearing thousands of issues a day.

The problem it solves is the one agents created. A coding agent checks its work by running code and reading the result, over and over, and doing that against the live cloud costs money and seconds; locally it costs neither. It began as a weekend experiment in which Amazon's Kiro agent produced a passable version of the SQS queue service in 15 minutes. The limits are real — an IDC analyst told InfoWorld it does not simulate throttling, outages or regional feature differences and does not sandbox the code it runs — and AWS publishes the known gaps in a GitHub folder literally called "things-that-do-not-work". Notable: AWS released this free and public-domain while LocalStack spent the year putting its free tier behind an account and a noncommercial licence.

Plurai raised $13 million to test agents inside a digital twin of your company. Team8 led the seed, with U&I, Velocity and Mercer participating; founders Ilan Kadar and Elad Levi came from Nexar, where Kadar worked with Waymo on simulating and evaluating its driving software, TechCrunch reported. The product copies a company's knowledge base and tools, maps every flow an agent can take, and generates synthetic edge cases to break it before a customer does.

The demand is easy to justify. LiveBench 2.0 has the best models at 82.9 and 75.6 out of 100 on long agentic tasks, and a simulated corporate environment built by EON put the strongest at 76.8% accuracy. Demos are cheap; the distance between a demo and a system you can leave alone is the product, and that distance is what this round is buying.

Which of today's four — a model that only decides, an emulator nobody read, a $1.5 trillion book, or a decade of cures — changes your week? Write to us via our contact page.