Daily AI digest, November 10

Felix Laurent 11 Nov 2026 6 min read

News

The biggest AI news of November 10 came from Anthropic and Google. Anthropic launched office tools that sent Microsoft shares down about 3%, then set its IPO range at $70 to $78 a share. Google opened Gemini 4 Carbon to paying developers and released a new robotics model. OpenAI brought verified businesses into ChatGPT, and CoreWeave's backlog reached $132.5B. Below are the big stories of Tuesday, November 10, plus the day's rumours.

Anthropic moves into office software

Anthropic launched Claude Sheets at an event in San Francisco. The beta builds and edits spreadsheets with formulas, multiple tabs and shared editing, and it exports to Excel. Claude Whiteboards also arrived in beta for paid users.

Claude Meetings is a note-taking agent for Team and Enterprise plans, starting with Google Meet. Claude Dashboards became Claude Data, Engadget reports. Microsoft shares dropped about 3% during the day after the announcement and closed down 2.2%. We checked the numbers and what Sheets can do here.

Anthropic sets its IPO range at $70 to $78

Anthropic's amended S-1 offers 825 million Class A shares at $70 to $78 each, which could raise up to about $64.4B. It expects to price the offering on November 18 and start trading on November 19, CNBC reports. The implied value depends on the share count: about $1.02 trillion on 13.09 billion shares, and up to $1.4 trillion in Yahoo Finance's report.

A day earlier, a range of $70 to $80 reached the banks and was withdrawn hours later, as we explained here. TSMC's board approved buying up to $500M of Anthropic stock at the IPO price, and up to $500M of OpenAI preferred stock.

The filings also show a crack in Anthropic's compute plans. Anthropic and Lambda can now each cancel a deal worth about $34.9B, because Lambda missed a deadline to secure a facility lease. Anthropic says it does not plan to terminate "at this juncture", Crypto Briefing reports.

Google ships Gemini 4 Carbon and a new robotics model

Google made Gemini 4 Carbon available to paying developers and began rolling it out to AI Ultra subscribers. The API costs $2 per million input tokens and $10 per million output tokens through November, then both rates double. Our breakdown of the price is here. Alongside it, Google added /swarm to Antigravity, a preview that runs teams of agents on a shared message board; details are here.

DeepMind released Gemini Robotics 2.5, which sorted car parts with 96% success in a pilot at Mercedes-Benz plants. We covered the results here.

DeepMind also introduced AlphaProof Code, whose agents formally verified key security properties of Signal's protocol code. The work took four weeks and produced over six million lines of Lean, with humans reviewing about 1%. More in our article.

OpenAI brings businesses into ChatGPT

OpenAI launched Businesses in ChatGPT in beta. Verified companies get a profile they can update, plus data on how often ChatGPT mentions them. Eligible US e-commerce and online-service firms can also run Business Agents, which must disclose that they are AI. Our breakdown is here.

OpenAI also hired Francis Davidson, the founder of AI travel startup Odessia, in what TechCrunch describes as an apparent acqui-hire. Separately, it described Control Room, a service that can stop an internal model's output within 35 seconds. The service denies inference if it goes down, Interesting Engineering reports.

CoreWeave beats estimates as its losses grow

CoreWeave reported third-quarter revenue of $3.66B, up 168% from a year earlier and above the $3.55B analysts expected. Its backlog rose to $132.5B from $104B in June, and it raised its 2026 revenue forecast to between $13.0B and $13.3B.

Its net loss widened to $709M from $110M, and debt reached about $47B, CNBC reports. Management said it does not plan parent-level unsecured debt through 2027 at current yields. Those yields follow the bond selloff we covered here.

Breaches spread across AI developer platforms

Railway published its incident report on an October attack that led to about $740,000 in fraudulent AI usage. Attackers could have decrypted environment variables in up to 1.85% of workspaces, about 100,000. Tokens for 1,852 users were exposed, BleepingComputer reports.

Notion said the Fireworks AI breach, which we reported on Monday, may have exposed some Notion AI chat transcripts. Under 1% of workspaces could be affected, and Notion has disconnected from Fireworks.

Smaller updates

Sixteen smaller stories came out on Monday evening and Tuesday.

Nvidia shares closed down 5.47% on Tuesday. The independent directors retained Wachtell Lipton for the investigation we covered here.
Australia's prime minister announced a new AI safety regulator with the power to levy significant penalties, Nikkei Asia reports.
Google launched Gemini for healthcare in preview in the US, with Northwell, HCA and Optum among its early customers, Fortune reports.
ElevenLabs released Scribe v3, which makes 27% fewer word errors than Scribe v2 across 102 languages. Details are here.
MiniMax released H3-Image, a 16-megapixel open-weight model whose license excludes the US, EU, UK and South Korea. More in our article.
Kora Systems unveiled Kora OS, a free Linux desktop built around AI agents and due in 2027. We covered it here.
Google AI Edge Gallery added offline speech and XR Language Teacher, a voice tutor on Gemma 4 that explains what the phone camera sees. Details are here.
NVIDIA released two synthetic Japanese datasets, including 9 million fictional personas with financial details. Details are here.
NVIDIA also presented GR00T 2, a robot model that reached 71% average success with one demo per task. Weights are due in December.
Meituan published the open weights of LongCat-2.5 under an MIT license, a week after the model's launch, VentureBeat reports.
Amazon made the Bedrock AgentCore Web Search API generally available at $7 per 1,000 queries.
Graphon Labs, which builds memory for AI agents, raised a $35M Series A led by RRE Ventures, The Next Web reports.
Users of Claude Code and Codex report new blocks on reverse engineering. An OpenAI engineer called them "mostly accidental", 404 Media reports.
Synaptics' patent claims ended work on Vino, an open-source DisplayLink driver built with LLM help, LWN reports.
FraudBench, a new academic test of 20 models, found fraud rates of 3.0% when advising and 21.4% when acting as agents.
We tested Claude Fable 5.2 in two three.js scenes and compared it with GPT-6 Astra and Gemini 4 Argon.

Three of the sixteen are open models or datasets.

Rumours

Seven more stories circulated on November 10, but we could not read the primary report or confirm them with the companies.

OpenAI's third-quarter revenue reportedly reached about $10B, up roughly 50% from the second quarter, The Information reports.
A Point72 unit has reportedly sued a fired analyst, alleging Claude and Codex were used to take source code, Bloomberg reports. These are allegations.
Apple is reportedly preparing Harbor, a Mac app that serves open-weight models across Macs linked by Thunderbolt 5, according to a leaker cited by 9to5Mac.
The Commerce Department has reportedly delayed Entity List additions for two months, leaving Megaspeed off the list, the Financial Times reports.
Sakana AI has reportedly hired Goldman Sachs for an IPO as soon as autumn 2027, Nikkei says. The company says no decision has been made; more here.
Elon Musk reportedly said Grok 4.8 has finished pre-training and could reach some users this week. No benchmarks or release date were published.
Rackspace and Together AI have reportedly signed a deal worth about $380M for Nvidia GB300 clusters in 2027. We could not read the release.

Treat all seven as unconfirmed until the companies announce them.

What the day was about

Tuesday showed the labs reaching deeper into everyday business software. Anthropic went after spreadsheets and meetings, and OpenAI gave companies profiles and agents inside ChatGPT. Microsoft's 3% drop shows investors take that competition seriously.

The money behind it is getting a public price. Anthropic now has a range and a listing date, and CoreWeave's backlog grew even as its losses did. On November 10 the question for AI companies moved closer to what public markets will pay for all this compute.