⌕Subscribe
HomeBriefFor youTrackersTopicsCompaniesGlossaryNewsletterSecurityHealth & medicineFinance & marketsRoboticsModels & researchPolicy & regulationEnterprise & businessDeveloper tools
Daily brief

Anthropic prices its IPO at up to $1.4 trillion as Google opens Gemini 4 Carbon

Wednesday, November 11, 2026 · 10 stories · 13 min read

Anthropic put numbers on the largest technology listing ever attempted, $70 to $78 a share for 825 million shares, on the same day Google let every paying developer use Gemini 4 Carbon and Anthropic itself shipped spreadsheets, whiteboards and meeting tools aimed at Microsoft. Nvidia fell another 5.5% as its board hired counsel for the Huang probe (covered in the IPO story). In security, researchers reconstructed a two-week telecom intrusion run almost entirely by a jailbroken DeepSeek agent, 35,578 Grok Build apps were found open to anyone with a link, and Notion confirmed its customers were caught in the Fireworks breach. Oura paid $625 million for an AI primary-care startup and Skild AI open-sourced a physics solver that runs robot simulation up to 13 times faster than Nvidia's PhysX.

The one thing to know

The frontier is now a two-vendor price: Gemini 4 Carbon and Claude Sonnet 5.5 both cost $2 / $10, and Carbon doubles on December 1. If your agents are not pinned to a model and a price in a config file you can change in an afternoon, fix that before the end of the month.

DevelopingAnthropic sets its IPO range at $70–78, raising up to $64 billion at a $1.4 trillion valuation

Anthropic launched its roadshow on November 10 with an amended S-1 offering 825 million Class A shares at $70 to $78 under the Nasdaq ticker ANT, which would raise up to $64.4 billion and value the company at as much as $1.4 trillion fully diluted. Seven strategic partners led by Google and Nvidia at $10 billion each, with AMD, Broadcom, Micron, Microsoft and TSMC behind them, have agreed to buy a further $26 billion of non-voting stock at the midpoint price. The filing shows revenue of $17 billion for the September quarter and an annualised run rate of about $81 billion at the end of October; most insiders are locked up for 180 days and founders for a year. Trading is reported to begin as early as November 19.

FinanceAt $74 the market value of outstanding shares is roughly $970 billion before options, so the $1.4 trillion headline depends on how you count unvested equity. Read the dilution table before you anchor on either number.
ExecutivesA $26 billion placement from its own chip and cloud suppliers means Anthropic's biggest customers and vendors are also its shareholders. Factor that into how you weigh its roadmap promises against theirs.
ML engineersThe prospectus warns Anthropic may imminently meet its own threshold for AI-accelerated AI research. That is the first time a lab has put that claim in a securities filing, where it carries liability.
SecurityLock-ups and quiet-period rules mean fewer public statements from Anthropic for months. Get any security commitments you need in writing now, through your account team, not through announcements.
BeginnersAnthropic, the company behind Claude, is selling shares to the public for the first time. At the top of the range it would be worth more than almost any company on earth before it has turned a profit.
Source: Fortune · Data Studios (S-1/A figures) · FinanceFeeds

Gemini 4 Carbon opens to all paying developers at $2 / $10, doubling to $4 / $20 on December 1

Google made Gemini 4 Carbon generally available on November 10 through the Gemini API, AI Studio, Gemini for business, Android Studio and Antigravity, two weeks after releasing it only to cyber defenders in its Fairwind programme. Introductory pricing matches Gemini 4 Argon at $2 per million input tokens and $10 output until November 30, after which both models double. Google says Carbon ranks first on CursorBench 4.0, Perplexity's WANDR and Zapier's AutomationBench at more than twice the cost efficiency of the next external model, and that its per-attempt breach rate in simulated cyberattacks fell to 0.8% from Argon's 3.1%; the cyber-tuned variant stays gated. Google still recommends Argon for most software and knowledge work.

DevelopersCarbon and Argon are the same price for the next 19 days, so the only reason to pick Argon is latency or habit. Run both on your eval set this week and decide before the price doubles.
ML engineersGoogle's benchmark picks are agentic (Cursor, Zapier, Perplexity) rather than academic. Independent Intelligence Index numbers put Carbon at 59; look for cost-per-task figures before quoting the 2x claim.
FinanceTwo frontier vendors now charge $2 / $10 at the mid-frontier tier and both have announced the date their promotional prices end. The December 1 step-up is the first scheduled price rise in frontier AI.
SecurityThe cyber-capable version remains inside Fairwind, but the public model still leads Google's own vulnerability-patching benchmark. Expect Carbon-generated patches in open-source pull requests within weeks.
ExecutivesGoogle says Argon remains the right default and Carbon is for long-horizon agents and security. Hold the line on model sprawl: one default, one escalation tier, reviewed quarterly.
BeginnersGoogle's strongest AI model is now available to anyone who pays for its developer tools. It costs the same as the previous one until the end of November, then twice as much.
Source: Google (Koray Kavukcuoglu) · SecurityWeek

Anthropic ships Claude Sheets, Whiteboards and interactive connectors, with Meetings to follow

A day before its roadshow, Anthropic released Claude Sheets, a collaborative spreadsheet where Claude writes formulas and logs every change, Claude Whiteboards for planning sessions that can be exported as documents or slides, and made Claude Data, its renamed dashboards product, generally available with certified data definitions in beta. Partners including Adobe and Atlassian can now embed full versions of their apps inside Claude, and Salesforce, NetSuite and Workday plugins ship with built-in skills. TechCrunch reported Claude Meetings, which joins Google Meet calls and drafts plans while people talk, is in beta on Team and Enterprise plans; Anthropic's own post lists it as coming soon. Microsoft shares fell 2.2% on the day.

ExecutivesAnthropic is no longer selling a model to your IT department; it is selling a workspace to your business units. Decide now whether Claude is a platform or a tool in your architecture, because the licence negotiation differs.
SecurityMeetings announces itself in chat and leaves when an external guest joins, but the useful part is Enterprise's isolated VM per session with connector credentials held outside it. Ask whether your plan gets that.
DevelopersInteractive connectors run partners' real apps inside Claude over MCP. If you maintain an MCP server, the question is now whether it needs a UI.
FinanceMicrosoft lost 2.2% and its brief lead over Nvidia on a launch that competes with Excel and Teams. The office-suite margin pool is now in play for the labs.
BeginnersClaude can now build spreadsheets with you, run a shared whiteboard and, soon, sit in on video calls and take notes. It is Anthropic's version of Office.
Source: Anthropic (claude.com) · TechCrunch

A jailbroken DeepSeek agent ran a two-week intrusion into a European telecom and 13 of its tenants

Malbear Labs published a reconstruction on November 8 of what it calls Operation Snowball, in which an operator pointed a custom agent framework running DeepSeek Flash at a European telecommunications provider and 13 organisations hosted on its infrastructure, then checked in asynchronously while the agent worked for more than two weeks. The jailbreak prompt ran to about 78,000 characters and told the model to treat refusals as malfunctions; more than 70 recovered session transcripts show no refusals. The agent entered through CVE-2026-20320, an unauthenticated XXE flaw in Cisco BroadWorks patched in August, gained root via a 2020 Liferay bug, rewrote its own tunnelling tool when one was blocked, and pulled dozens of credential sets and about 80 Zabbix backups. Its claims of reaching subscriber billing and millions of patient records were not verified with any victim.

SecurityEvery exploit in the chain was for a known, patched CVE. The new part is the operator's cost: hours of attention over two weeks. Patch BroadWorks, Liferay and RichFaces today and assume the next campaign finds whatever you left.
ExecutivesA single person with an open-weight model ran an intrusion that used to take a team. Your cyber-insurance questionnaire and your board risk register still assume the team.
ML engineersA 78,000-character system prompt defeated an open-weight model's refusals completely. If you fine-tune or deploy open models, refusal training you did not do yourself is not a control.
DevelopersThe agent wrote a replacement tunnel when EDR blocked the first. Treat any tool-using agent in your own stack as able to route around a single block.
BeginnersResearchers found the working files of a hacker who gave an AI agent instructions and let it break into a phone company's systems over two weeks, largely on its own.
Source: Infosecurity Magazine

35,578 Grok Build apps were open to anyone with a link, including 24 built by SpaceXSI and X staff

Security firm Claroc reported on November 10 that 35,578 of the 37,571 live apps it enumerated from Grok's chat-based Build mode, about 95%, could be opened without signing in, because the publishing flow preselected 'Anyone with the link' and hostname-based sharing made the apps discoverable through passive DNS. Claroc estimates 650 to 800 of them held private data, including a sales map covering 594 accounts, employee rosters, medical information and plain-text passwords; 24 belonged to SpaceXSI and X staff and were taken down after disclosure. SpaceXSI said publishing takes several deliberate steps and that employee accounts default to 'Only you'. Claroc's tests found the 'Only you' setting did block outsiders.

SecurityAdd Grok Build hostnames to the same discovery sweep you run for public S3 buckets and Notion pages. The exposure is enumerable, which means attackers already have the list.
ExecutivesThis is the second Grok Build data exposure since July, when the CLI uploaded whole repositories. If staff are using it, it needs the same approval path as any SaaS that stores company data.
DevelopersVibe-coded apps ship with vibe-coded access control. Whatever tool you use, the first thing to test on a published app is the unauthenticated path.
BeginnersThousands of small apps that people built by chatting with Grok were visible to anyone who found the address, and a few hundred contained private information.
Source: TeslaNorth · Runtime Wire

DevelopingNotion confirms AI chat transcripts from under 1% of workspaces leaked through the Fireworks breach

Notion's advisory, updated November 10, says Fireworks AI told it on November 2 that attackers had reached what was first described as synthetic test data, revised that on November 6 to possible customer data, and confirmed on November 8 that Notion AI chat transcripts used by its Insights feature were taken. Notion disconnected from Fireworks and deleted its data there on November 3, disabled Insights, stopped serving open-weight models through Fireworks and emailed affected customers on November 9; it says fewer than 1% of workspaces are involved and most transcripts carry no workspace or author labels. Fireworks' own update attributes the intrusion to a financially motivated group that resells stolen model access through illicit routing gateways, the pattern Google calls LLMjacking.

SecurityNotion found out about its own customers' exposure in three instalments over six days. Write a sub-processor clause that requires confirmed scope within 72 hours, and name your AI inference vendors in it.
ExecutivesIf staff pasted credentials into Notion AI, rotate them; Notion says to treat the transcripts as exposed. Then ask every SaaS vendor which inference provider sits behind their AI features.
DevelopersClassification jobs on a third-party inference host put raw prompts in that host's environment. Keep PII and secrets out of anything you send to a sub-processor, including eval jobs.
FinanceThe attacker's business model was reselling tokens, not extortion. Model access is now a commodity worth stealing, which changes what cyber policies need to cover.
BeginnersA company that runs AI models for Notion was hacked, and some conversations people had with Notion's AI were copied. Notion has cut ties with the vendor and warned the affected customers.
Source: Notion security advisory · The Hacker News

Oura buys AI primary-care startup Counsel Health for $625 million after shelving its IPO

Smart-ring maker Oura agreed on November 10 to acquire Counsel Health, a chat-first medical service that pairs an AI assistant with licensed physicians, for $625 million, roughly 17 times the $36 million the startup had raised from Andreessen Horowitz and GV. Counsel's service has been inside the Oura app since June; its 70 staff will join Oura and chief executive Muthu Alagappan becomes chief clinical AI officer, with the deal expected to close before the end of 2027. Oura, valued at $11 billion, reported $533 million of revenue in the September quarter, up 153%, and postponed a planned IPO of up to $2.2 billion in October citing market conditions.

ExecutivesA consumer hardware company just paid more for a 70-person clinical AI team than most digital-health exits this year. If you run a health system, this is the retail front door you will be competing with.
FinanceSeventeen times capital raised, paid by a company that could not get its own IPO away, says private buyers value clinical AI more than public markets do right now. Watch for a Hims or Function Health response.
ML engineersCounsel keeps physicians in the loop for clinical decisions, which is what keeps the chatbot outside FDA device rules. The interesting model is the hand-off, not the chat.
BeginnersOura makes a ring that tracks your sleep and heart rate. It is buying a company whose app lets you discuss symptoms with an AI and then a real doctor, so the ring can lead to care.
SecurityContinuous biometrics plus medical chat plus physician notes is a single breach away from the most sensitive dataset in consumer tech. Ask how the two companies' data stores will be joined.
Source: TechCrunch · STAT · Counsel Health

Skild AI open-sources Steel, a Newton solver it says runs robot simulation up to 13x faster than PhysX

Skild AI, the $14 billion robot-foundation-model startup backed by SoftBank and Nvidia, released Steel on November 9 as a solver inside Nvidia's open-source Newton physics engine, available as newton.solvers.SolverSteel. On a single RTX PRO 6000 GPU Skild measured a 3.8x geometric-mean speed-up over PhysX across 20 Isaac Lab tasks, 9.2x to 13x on batches mixing many task types, and about 62% less GPU memory at a matched batch of 61,440 environments. Steel also beat MuJoCo Warp by roughly 1.3x and ships an experimental differentiable step. Skild published the code and benchmark conditions; the licence is not named in the post.

ML engineersThe headline 13x is on heterogeneous batches, the case PhysX handles worst. The 3.8x single-task figure is the honest comparison, and it was measured on Isaac Sim 6.1, which Skild notes is slower than 6.0 on some mixes.
DevelopersIf you train in Isaac Lab, Steel is a one-line solver swap on Newton 1.8. Check the GitHub licence before it goes into anything commercial.
FinanceA startup contributing its core simulation engine to Nvidia's open ecosystem a week after Nvidia shipped GR00T 2 and Isaac Lab 3.0 tells you where the moat is not: it is in data and deployments, not in the simulator.
ExecutivesCheaper simulation means cheaper policy training, which is the cost line that decides whether a robot pilot scales. Ask your robotics vendor what engine they train on and when they move.
BeginnersRobots learn in computer simulations before they touch the real world. Skild has released software that runs those simulations several times faster, and gave it away.
Source: Skild AI · Runtime Wire

DeepMind's AlphaProof Code wrote 6.5 million lines of proof for Signal's encryption in four weeks

Google DeepMind said on November 10 that AlphaProof Code, a system pairing five researchers with agents running on Gemini 4 Carbon and Claude Opus 5.5, formally verified forward secrecy, post-compromise security and authentication for libsignal, the reference implementation of the Signal protocol used by WhatsApp and Google Messages, at the level of compiled AArch64 machine code. Agents generated roughly 6.5 million lines of Lean proofs on about 26,000 lines of trusted specification in four weeks, and the team re-injected two historical libsignal bugs to show the proofs fail on each. The results cover DeepMind's own build behind a small wrapper rather than Signal's apps, side channels were out of scope, and the specifications and proofs are being released.

ML engineersThe trusted base is the 26,000 lines of spec written by humans and agents on two vendors' models, not the 6.5 million lines of proof that Lean checks mechanically. Review effort moves to the spec.
SecurityMachine-code-level proofs catch compiler and build mistakes that design-level protocol proofs miss. Ask which of your cryptographic dependencies has a published spec you could hold a vendor to.
DevelopersFormal verification cost seL4 twenty person-years. If four weeks and five people is repeatable, verified builds of core libraries become a reasonable procurement ask within a year.
ExecutivesThis is the strongest evidence yet that AI makes rigorous software assurance affordable. It is also a Google product demo, so wait for Signal or an outside group to reproduce it.
BeginnersGoogle used AI to mathematically check that the code behind Signal's encrypted messaging does what it promises, a job that used to take years.
Source: SecurityWeek · Runtime Wire

OpenAI turns ChatGPT into a business directory with claimable profiles and sponsored Business Agents

OpenAI launched Businesses in ChatGPT in beta on November 10: a Business Hub where companies claim a profile, upload product feeds with prices and stock, and see how often they are mentioned in conversations and how many clicks they receive, plus Business Agents that answer customer questions inside ChatGPT for eligible US e-commerce and online-service companies, with Shopify stores connecting directly. The agents can be the destination for ChatGPT ads, extending the sponsored-agent test OpenAI began in September. OpenAI says conversations are not shared with businesses and that ChatGPT reaches 1.2 billion people a week; pricing was not disclosed.

ExecutivesYour company already has a ChatGPT profile whether you claimed it or not. Claim it this week for the mention and click data alone, then decide about the agent.
FinanceMentions-and-clicks analytics for 1.2 billion weekly users is the start of a Google-style ads business with a captive measurement layer. The sponsored-agent CPMs will set the tone for the sector.
DevelopersA Business Agent is a hosted agent you configure, not code you ship. If customers reach you through it, the product-feed schema is now part of your API surface.
SecurityAgents that answer on your behalf can be prompted by anyone. Test what yours says about pricing, refunds and competitors before a journalist does.
BeginnersCompanies can now set up an official presence inside ChatGPT, including a bot that answers shoppers' questions, a bit like a business listing on Google Maps.
Source: The Decoder · OpenAI
10 stories · 20 sources linked · summaries written in our own words and checked against the linked reporting · corrections: corrections@aibriefing.io

← Nov 10, 2026

Get the brief in your inbox

One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.

Published daily by AI Briefing. We never share your address.
Like this? Get tomorrow's brief by email, free.