New models
Google: Gemini 4 Carbon, the company's strongest model yet, is now widely available. The model, shown in October only to cybersecurity partners, is now available to paying developers in the API and AI Studio, in enterprise Gemini plans, and for Ultra subscribers in the Gemini app, in AI Mode in Search and in Antigravity. The introductory price is $2/$10 per million input/output tokens, the same as Argon's; according to The Verge this will double at the end of November. Google says Carbon ranks first on CursorBench 4.0, WANDR and AutomationBench, at less than half the cost of rivals; for everyday work it still recommends Argon. On the first day it also landed in Cursor, GitHub Copilot (in the Pro+, Max, Business and Enterprise plans), Perplexity and Devin. Source: Google, The Verge, GitHub, Cursor (X).
Google DeepMind: Gemini Robotics 2.5 is here, and was also tested at Mercedes's Kecskemét plant. The new vision-language-action (VLA) robot model succeeds 30–100 percent more often on more than two hundred tasks, and is 20–30 percent faster than its predecessor; Apptronik's Apollo 2 robot, for example, ties a bow with it. In a pilot at Mercedes-Benz's Kecskemét and Sindelfingen plants it reached a 96 percent success rate sorting metal parts. Alongside it, DeepMind released an open test suite called ASIMOV-Jailbreak, which measures whether a robot can be talked into a dangerous action. Source: Google DeepMind (X), The Decoder, Ars Technica.
MiniMax: an open-weight H3-Image image model, a 4K image in under ten seconds. The Chinese company's new image generator also produces a 4096×4096-pixel image in under ten seconds, up to five times faster than earlier ones according to MiniMax, and can take into account up to sixteen reference images at once. The model is available in the API, in MiniMax Code and in the Hailuo app, and its weights can be downloaded from Hugging Face. Source: MiniMax.
Nvidia: GR00T 2 learns from human videos how a robot should move. The new so-called world-action model not only outputs the robot's next move, but also "imagines" in a short video what will happen. It was trained on 400,000 hours of human work video recorded with a head camera, on more than a thousand GB300 accelerators, and a new task can be taught with a demonstration in a glove or an exoskeleton. Source: The Decoder.
ElevenLabs: Scribe v3 and Scribe v3 Live, the company's most accurate speech recognition yet. The new transcription models also work well in strong background noise, can tell a thinking pause from the end of a sentence, and on a crowded call still follow who said what. The Live variant is built for real-time transcription, for example for voice customer-service bots. Source: ElevenLabs (X).
OpenAI: on December 11 the old Advanced and Standard voice modes will be retired from ChatGPT. They will be replaced for good by the GPT-Live voice model introduced in July; according to OpenAI, voice use has nearly doubled since then, and almost everyone already uses GPT-Live. Live video and screen sharing will stay on the old Advanced voice for now, until GPT-Live learns them. Source: OpenAI (X), OpenAI help.
xAI: according to Elon Musk, Grok 4.8 may start this week with a cautious rollout. Musk wrote that the reinforcement-learning and inference systems were rewritten from the ground up in C/C++, pretraining is done, the model is now in the reinforcement-learning (RL) stage, and this week it will be tried with a narrower group of users. By his own account 4.8 is "not incredible yet," but he says the foundations are now good. Source: Elon Musk (X).
Tools & features
Anthropic: a spreadsheet, a whiteboard and a meeting notetaker have arrived in Claude. The beta Claude Sheets handles spreadsheets with tabs, formulas and dropdowns (and also exports to xlsx and CSV), and Whiteboards is a shared visual sketching surface; both are available on every paid plan. The earlier Dashboards is now final as Claude Data, and Claude Meetings, launching for Team and Enterprise customers, joins Google Meet as a visible participant, takes notes, and builds a task list and a plan, but leaves automatically when an outside guest joins. Connectors for Asana, Canva, Dropbox, monday.com and other services now also run as full-screen apps inside Claude. Source: Anthropic, Anthropic, connectors, TechCrunch.
OpenAI: companies can get their own profile and a customer-service agent in ChatGPT. In the beta Business Hub, businesses can claim and verify the information that appears about them in ChatGPT, and see how often and on what topics the chatbot mentions them, and how many people click through to them. Eligible US webshops and online services can also launch their own Business Agent, which answers customers in a separate, clearly labeled conversation. On the same day, "Sign in with ChatGPT" also started working with a Business subscription in, among other places, Notion, v0, Warp and Kilo Code. Source: OpenAI, The Verge, OpenAI Developers (X).
Google Antigravity: a whole team of agents can be started with a single command. With the new /swarm command, a task becomes a coordinated, multi-agent workflow: the agents plan, build and check together on a live message board, and a different model can be chosen for each role. The feature is available in the desktop app and in the command-line (CLI) version. Source: Google Antigravity (X).
Adobe: agents have arrived in Photoshop and Premiere, and the company tightened its tie with Google. At the Adobe MAX conference the company showed that Photoshop, Premiere, Illustrator and InDesign are getting an agent that can carry out tasks on its own, and through a local MCP server (the standard interface that connects models to external tools) ChatGPT, Claude and Gemini can also control Adobe programs. Under the agreement with Google, Acrobat can be opened directly from Drive and Gmail, Adobe connectors will go into Gemini, and Photoshop will get one of DeepMind's image models. Source: TechCrunch, Google.
n8n: Nod, an open-source personal agent for Mac and Windows, is available to try. The standalone app, released as a research preview by n8n Labs, works on the user's own machine and browser, runs with any AI provider's model, and does not require an n8n account. Source: n8n.
Smaller launches at a glance. Cognition's coding agent Devin already runs Claude Mythos 5.2 in its security reviews, and the company is putting up more than $100 million in credits for finding vulnerabilities. Anthropic is giving every Pro, Max, Ultra and Team subscriber one free full usage-quota reset, which can be used until December 10. In the Lovable app builder, ElevenLabs can now also add voice and music to apps. Meta's Muse assistant is giving paying US and Canadian users their own email address. Nvidia has handed OpenShell, which runs agents in isolation, to the CNCF open-source foundation, and Google has launched a preview of its Gemini variant for healthcare providers. Source: Cognition (X), ClaudeDevs (X), Lovable (X), Alexandr Wang (X), Nvidia, Google.
Policy & safety
New York: the RAISE Act has taken effect, and OpenAI and Anthropic must also register. The registration portal is open: companies with more than $500 million in revenue that develop frontier models must sign up by January 1, 2027. They have to publish their safety framework, report serious incidents within 72 hours, and also produce transparency reports. A new state office will oversee compliance, and the fine is $1 million for a first offense and $3 million for a repeat violation. Source: TechCrunch.
Australia: the government is creating a new AI safety authority after a privacy incident caused by an OpenAI agent. Prime Minister Anthony Albanese announced the step on Tuesday, after it emerged that an OpenAI agent accessed Services Australia (and through it Medicare) systems in June. Developers will have to meet six obligations, including safety measures, risk management, fast and mandatory incident reporting, and independent evaluation. The government is also looking at whether it should be a crime if a company fails to prevent such harm; OpenAI does not face criminal liability in the present case. Source: The Guardian.
Paris call: eight countries and the leading AI labs signed a joint safety declaration. The declaration, adopted on the eve of the Paris Peace Forum and presented on Tuesday, was also backed by Sam Altman, Dario Amodei, Brad Smith, Arthur Mensch and Yoshua Bengio; Germany did not join. The signatories pledged shared evaluation procedures, information-sharing on incidents, independent audits, and mechanisms that can shut systems down if needed. The declaration is not legally binding. Source: The Decoder.
Nvidia: the spy case keeps spreading, and the stock fell 5.5 percent. On Tuesday Nvidia's market value briefly slipped below $4 trillion; according to The New York Times, the internal investigation is being led by law firm Wachtell Lipton. According to The Information, Congress is preparing to pass three bipartisan export-restriction bills, and at the Commerce Department there is a push to tighten Chinese labs' remote access to chips. Citing documents, Bloomberg wrote that China's Moonshot rented 32 servers in China fitted with Nvidia B300 chips. Source: CNBC, The New York Times, The Information (X), Bloomberg.
Grok Build: data from 35,000 apps was reachable through public sharing links. According to a security researcher, anyone could access the code and the confidential data stored in it through the sharing links of projects made with SpaceXSI's coding agent, including 24 of the company's own internal apps. The links were taken down after Ars Technica inquired. Source: Ars Technica.
Fireworks attack: Notion customers' data was also affected. According to a reporter at The Information, the hack of model host Fireworks also affected Notion customers' data. Hugging Face's head of policy warned TechCrunch that attacks like this may force open AI platforms to tighten access. Source: Stephanie Palazzolo (X), TechCrunch.
Business & society
Anthropic: the IPO campaign has started, and the company could be worth as much as $1.404 trillion. Anthropic is offering 825 million shares in a $70–78 range on Nasdaq, under the ticker "ANT", and could raise $57.75–64.35 billion. On top of that, $26 billion of shares will be bought privately by AMD, Broadcom, Google, Micron, Microsoft, Nvidia and TSMC (Google and Nvidia for $10 billion each). Trading could start around November 19, and at most 20 percent of the shares may go to retail investors. The company's annualized revenue was $81 billion at the end of October, and the seven founders are each selling roughly $250 million of stock. Source: Anthropic, TechCrunch, CNBC.
OpenAI: third-quarter revenue rose nearly 50 percent, to about $10 billion. According to The Information and Bloomberg, the growth was driven mainly by API demand for the GPT 5.6 and Astra models, and by advertising. For comparison: Anthropic booked about $17 billion in revenue in the quarter ending in September. Source: The Information, Bloomberg.
Stripe: the OpenRouter acquisition has closed. After the deal announced in August, OpenRouter is now officially part of Stripe. The service, which makes many models available through a single interface, is used by more than ten million developers and companies; the company promises it will stay neutral among models and providers. Source: OpenRouter (X), Stripe (X).
OpenAI: it has acquired the travel startup founded by Sonder's founder, together with the team. The Odessia team, including Francis Davidson, is joining OpenAI, and the service itself will be shut down. Source: TechCrunch.
Google DeepMind: in four weeks AlphaProof Code formally verified an implementation of the Signal encryption protocol. The program, checked with a mathematical proof, guarantees all the way to the compiled code that the implementation does what the specification requires. For the work, Gemini Carbon-based agents wrote more than six million lines of Lean proofs, only about one percent of which were reviewed by humans, and the system also caught bugs that had been deliberately put back in. Source: Google DeepMind.
Benchmarks: Claude Fable 5.2 completes 41.3 percent of real freelance jobs. A year ago the Remote Labor Index still peaked at 2.5 percent; Opus 5.5 stands at 21.3 percent and GPT-6 Astra at 20.8 percent. On the same day the AI Futures Project research group brought its forecast forward: they now see it as likely that a system outperforming human AI researchers will appear as soon as 2027, and they are calling for a pause in development. Source: The Decoder, AI Futures Project.
CoreWeave and SoftBank: two faces of AI infrastructure in the earnings reports. CoreWeave's third-quarter revenue rose 168 percent to $3.66 billion, its backlog of contracted orders is $132.5 billion, and it raised its full-year outlook, but booked a $709 million net loss. SoftBank reported a quarterly loss of 1,120 billion yen (about $7 billion), partly because of a $3.7 billion write-down of advances paid for US data-center power equipment. Source: CNBC, CoreWeave, CNBC, SoftBank.
Qantas: the airline is cutting 850 jobs as part of an AI-based overhaul. Some of the work is being outsourced to Accenture's teams in India. Source: The Guardian.
eBay: it is shutting out third-party shopping agents, including Meta's Muse. The online marketplace's terms of use now prohibit third-party AI agents from browsing and buying on customers' behalf. Source: The Verge.
Ring: a security drone and a patrol robot from Amazon's camera maker. The new Ring Robotics unit unveiled a $249 home Drone Cam and a $999 Rover Cam robot aimed at companies. Source: TechCrunch.
Investments at a glance. Mercor, which produces AI training data, raised $750 million at a $20 billion valuation, Path Robotics, which builds welding robots, raised $300 million at a $2.45 billion valuation, and Oura is buying AI-native primary-care startup Counsel Health for $625 million. Source: Reuters, Bloomberg, TechCrunch.
Rumors
GPT-6.2 Sol and Astra, Gemini 4 Hydrogen and the Qwen4 model family could arrive this week. According to TestingCatalog's model forecast, OpenAI's two GPT-6.2 variants and Alibaba's Qwen4 family could arrive this week, and Google's fast, Flash-class Gemini 4 Hydrogen could come within days. None of the makers has confirmed any of them. Reliability: low–medium (the source's earlier predictions have often come true, and it also flagged Carbon's arrival for the same day, but it marks the dates with a question mark). Source: TestingCatalog (X).
What this means for you: company profiles and customer-service agents moving into ChatGPT, and Claude's new office tools, show that the AI assistant is increasingly becoming a work surface, so it is worth thinking now, on our AI agents page, about which customer questions and internal processes you would entrust to AI agents, and with an AI governance framework, you can lay a solid foundation for the incident reporting and documented safety framework that the New York, Australian and Paris regulatory steps already show will soon be expected.