Today in AI — November 10, 2026
Anthropic spent the day pricing the largest AI listing anyone has attempted, Google DeepMind spent it proving things — a cryptographic protocol all the way down to compiled code, a humanoid's hands in a Mercedes plant — and New York switched on the paperwork that will govern both. Here is Tuesday, in the order it matters.
Models & Research
Gemini Robotics 2.5 took a real factory job and kept 96% of it. Google DeepMind's new vision-language-action model ran Apptronik's Apollo 2 humanoid on active Mercedes-Benz lines in Kecskemét, Hungary and Sindelfingen, Germany, pulling metal brackets out of supply bins and packing them into sequencing containers at a 96% success rate, per DeepMind. Across 200 lab tasks — serving food with tongs, tying a bow on a gift box, pipetting liquid — it beats the previous generation by 30% to 100% on success and finishes 20% to 30% faster.
The honest number is the four points it misses. DeepMind director Kanishka Rao called closing that gap to human reliability a "steep road", and at factory volume 96% is still forty failures in a thousand parts, each one a technician's problem. Worth pairing with Nvidia's GR00T 2, which went the other way: learn from human video and skip the teleoperation. DeepMind's answer is a better model; Nvidia's is cheaper data. Both shipped the same afternoon.
DeepMind's agents formally verified Signal's cryptography, down to the compiled code. AlphaProof Code, a multi-model agentic framework, generated more than six million lines of Lean to prove that the implementation behind Signal's protocol keeps messages protected before and after a device is temporarily compromised, according to DeepMind. Specs were written with human experts reviewing every one; the agents then flagged just one percent of the codebase for a second human look and Lean machine-checked the rest.
The verification took four weeks. The same work has historically taken specialist teams years, and most protocol verification stops at the design level rather than the code actually running on phones — which is exactly where real vulnerabilities live. DeepMind also reintroduced two previously fixed bugs into an isolated copy; the system produced machine-checked proofs that both versions were insecure, which is the part that makes this evidence rather than a demo. It lands a day after the lab showed AI cutting the cost of RSA forgery, so the crypto story of the week now runs in both directions.
Agents now complete 41% of real freelance jobs at professional quality. Anthropic's Fable 5.2 leads Scale Labs and the Center for AI Safety's Remote Labor Index at 41.3%, per the leaderboard — against 21.25% for Opus 5.5, 20.83% for GPT-6 Astra, 17.08% for Gemini 4 Argon and 2.5% for the best system when the benchmark launched a year ago. The test is 240 real projects from 358 verified Upwork professionals worth about $144,000: $632.60 and roughly 29 human hours per job, judged by experts opening the submitted files in Blender, GIMP and Audacity.
The failures still cluster. Industrial product design and animation remain close to hopeless, and the typical failure is amateur work, a broken file or a self-contradicting design rather than a near miss. The same day, OpenAI and Upwork published their own study of where freelance demand is moving: of 63 work categories, 43 lost share on Upwork as their share of ChatGPT use grew, while AI-related postings bundled skills that rarely appeared together before. Read the two together and the story is not "agents take the jobs" — it is that the job listing itself is being rewritten.
On the reading list: "Big Math." Steven Strogatz and Alex Townsend spent five years on a book about the mathematics that only exists because machines can hold it — billions of variables, matrices nobody can picture — and Ars Technica's interview with them is the best hour of non-news reading published today. Their frame is 2,000 years old: the abacus for arithmetic, the fangcheng board for linear systems, now models that produce proofs faster than people can check them. "We're outsourcing our thinking to systems we don't fully understand," Strogatz says, and the interesting question is what mathematicians become when the answer arrives before the intuition.
Industry
Anthropic opened its roadshow at $70 to $78 a share. The company and existing holders plan to sell 825 million Class A shares, raising $57.8 billion to $64.4 billion at a fully diluted valuation of up to $1.405 trillion on Nasdaq under ANT, per the amended S-1. Bloomberg's Leslie Picker reported the deal is expected to price next Wednesday and begin trading Thursday.
The strategic money is the tell. Nvidia plans to put in about $10 billion, roughly a tenth of the new cash Anthropic expects to raise this month, the New York Times reported from the filing, and a $26 billion private placement at the IPO price brings in Google and AMD alongside it. The seven co-founders each sell 3,378,378 shares — near $250 million apiece at the midpoint — while keeping 51.1% of the votes through supervoting Class F stock. Board member Reed Hastings has signalled $300 million of his own. A trillion-dollar listing that leaves the founders in control and its chip supplier as a shareholder is not a normal IPO; it is a coalition being formalised.
SoftBank wrote off $3.7 billion of data centre equipment it had paid for in advance. The writedown sat inside a July-to-September quarter that produced a net loss of 1.12 trillion yen, about $7.1 billion, against a consensus forecast of a profit — the group's first loss in almost two years, Bloomberg reported. The payments were non-refundable — The Information put the committed total at $9.5 billion against equipment never delivered, so the $3.7 billion is gone rather than deferred. The pattern is worth naming: prepaying for GPUs and power gear turns a capex plan into credit exposure, and SoftBank just showed what the mark-down looks like when the plan changes.
Mercor raised a $750 million Series D, with Benchmark, General Catalyst and Nvidia among the investors, according to reports, in a deal that prices the AI training-data and expert-labour marketplace as one of the picks in the model-building supply chain. It is the least glamorous business in the boom and one of the most durable: every lab that ships a frontier model needs humans who can write the answers first.
Oura will pay $625 million for Counsel Health. The ring maker is buying the AI-native primary care company whose chat-plus-physician service has been inside the Oura app since June, TechCrunch reported. Revenue for the September quarter hit $533 million, up 153% year over year, and fiscal 2026 closed near $1.75 billion — which is why the company can spend $625 million on care delivery a few weeks after postponing its own $2.2 billion IPO.
Nvidia's board handed the Huang investigation to Wachtell. The independent directors retained Joon H. Kim, a former acting US attorney for the Southern District of New York, to run a "thorough and independent investigation" into whether company secrets or national security were compromised, CNBC reported. The stock fell 6% Tuesday after 4% on Monday, and a shareholder class action followed. The board's position is unchanged from Sunday: no reason to believe anything was compromised. The market is pricing the possibility that the board is wrong.
Policy
New York turned on the RAISE Act. The state opened its registration portal Tuesday; large frontier developers including OpenAI and Anthropic must sign up by January 1, 2027 and will answer to a new Office of Digital Innovation, Governance, Integrity, and Trust. Fines run to $1 million for a first violation and $3 million after that, TechCrunch reported.
Governor Kathy Hochul also launched the New York Center for AI Responsibility and Research at SUNY Binghamton, seeded with $55 million in donations and capital, to independently evaluate frontier systems and review companies' safety frameworks. Her line at the briefing was the quotable one: the tech moguls say "any collateral damage is worth it as long as we win the race to the AI cliff." New York now sits ahead of California's SB 53 and Illinois's SB 315 on enforcement rather than disclosure — a registration regime with a penalty schedule is harder to litigate away than a transparency rule.
Point72's Cubist sued a former star analyst over AI-assisted code theft. Cubist Systematic Strategies filed Monday in Connecticut state court in Stamford against Xin Jin, its former lead research analyst, alleging he used Anthropic's Claude and OpenAI's Codex to take proprietary source code on his way to a competitor; the firm says an internal investigation reached that conclusion and dismissed him in September, Bloomberg reported. This is the first prominent case that treats coding agents as exfiltration tooling rather than productivity software, and the discovery fight will be about agent logs — what the assistants were asked, what they read, and whether anyone thought to check before the laptop came back.
Tools
Anthropic turned Claude into an office suite. The app now carries Claude Sheets for spreadsheets with formulas and multiple tabs, Claude Whiteboards for brainstorming that Claude summarises, and Claude Data — the renamed Dashboards — wired into Snowflake and Salesforce, Anthropic announced. Interactive Connectors let partners including Adobe, Atlassian and Salesforce embed their apps inside Claude, Claude Meetings is "soon", and enterprise customers get audit logs, isolated virtual machines and a compliance API.
The connectors are the real move. Every feature above is table stakes against Microsoft and Google; being the surface where someone else's app runs is the position that compounds, and it is the same bet OpenAI made with its business profiles this morning. Two labs, one week, both deciding that the model is a commodity and the workspace is the product.
Adobe opened Creative Cloud to other companies' agents. At Max in Miami, Adobe previewed local Model Context Protocol integrations that let third-party agents from OpenAI, Google and Anthropic edit native project files in Photoshop, Illustrator, Premiere and After Effects, The Verge reported, alongside previews of always-on agents that run recurring work in the background and a Plan Mode for the Firefly assistant. Nothing ships until a private beta in early 2027.
The design choice is the interesting part: agents touch files locally, every layer and sequence stays editable, and the human can take over mid-task. Adobe is selling agent access to its format rather than ceding the workflow, which is the sane version of the deal every incumbent is now being offered.
eBay quietly closed the door on shopping agents. Meta's Muse stopped being able to pull listings and prices from the marketplace this weekend, and eBay's spokesperson told The Verge the company has "restricted agent access in certain cases, including unauthorized scraping and training on our data", The Verge reported. Amazon blocks external agents outright, Target blocks them intermittently, and Shopify has opened a billion listings to them.
The blocking is happening because almost nobody is actually shopping this way. Uber filled four rides through third-party agents last quarter and DoorDash took several hundred agent orders in a week against 993 million total. Retailers are building walls against a behaviour that has not arrived, which tells you they expect it to — and that the terms of agent access will be negotiated platform by platform rather than settled by a standard.
Which of today's three — verified code, a 96% robot, or a $1.4 trillion listing — actually changes your week? Write to us via our contact page.
