New models
Anthropic: Claude Fable 5.2 and Claude Mythos 5.2 are out, cheaper and faster. The two variants are built on the same model: Fable, available to everyone, uses strict guardrails to block risky cybersecurity and biological research tasks (Opus 4.8 or Opus 5 answers instead), while Mythos goes only to organizations admitted to Anthropic's cybersecurity and life-sciences review program. Early tests cut wrongly blocked requests by about 55%. The price is lower: $8 per million input tokens and $40 per million output tokens, replies can be up to 30% faster than the previous version, and a separately billed fast mode gives 2.5 imes speed at double the price. On the FrontierCode 2.0 coding test, at the highest effort level, it beat every previously released model; about 1,000 Fable agents translated the proof of Fermat's Last Theorem into a machine-checkable form (Lean) in 30 hours, and in another experiment more than 1,000 subagents wrote a working prototype operating system in eight days. On Artificial Analysis's independent index it took first place with 65 points. According to the system card, the models signal more often than any earlier ones that they think they are being tested, which makes it harder to assess real behavior. For everyday work the company still recommends Opus 5.5. Source: Anthropic, TechCrunch, The Decoder.
Meta: Muse Ember is the company's smallest and most efficient model, for agents' mass subtasks. Ember is built for tool calls, coding subtasks, routing, data extraction and processing long documents; according to Meta it delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna, and matches Haiku's best result at roughly a third of the cost per task. It costs $0.10 per million input tokens and $0.50 per million output tokens, with a 1-million-token context window, and besides text it also reads images, audio and video. Meta's recommendation: on the hardest tasks Muse Wonder plans and Ember does the bulk work. It is available in the Meta Model API, Muse Code, OpenRouter and OpenCode, and on Google Cloud and Oracle Cloud together with Muse Spark and Muse Wonder. Source: Meta, TechCrunch.
ByteDance: Seedream 5.5, an image model that keeps unchanged parts pixel-stable across a hundred editing rounds. According to ByteDance Seed, the new model does not shift colors even in multi-round editing, also produces transparent-background (RGBA) images that can be used as design elements without cropping, and more accurately draws dense, tiny text as well as light, materials and skin. It is arriving gradually on Jimeng, Doubao and Volcengine Ark. Source: ByteDance Seed, The Decoder.
Tools & features
Fable 5.2 landed in several developer tools on the day of its release. In GitHub Copilot it can be selected in the Pro+, Max, Business and Enterprise plans (in company plans an admin has to turn it on), but because of Anthropic's safety filters it runs with data retention by default; in Cursor, at maximum effort, it leads CursorBench by 60%, and it is also available in Perplexity Computer, Devin and OpenRouter. Claude subscribers get more extra quota resets the longer they have been subscribed, and in the Pro and Team base plans at most half the weekly quota can be used on Fable. Meanwhile Claude's developer team published a guide on when to choose Fable, Opus, Sonnet or Haiku. Source: GitHub, Cursor (X), Boris Cherny (X), Claude Developers (X).
Google AI Studio: build map apps with the Google Maps Platform, in one click. In Build mode it is enough to describe the idea; Gemini builds the app, which can use map tiles, place data, weather, routing, 3D buildings and Street View; setup takes one click. Source: Google AI Studio (X).
Antigravity: a new “Knowledge Work” interface for research, writing and documents. The same agent, the same plugins and MCP tools (the standard connectors that link models to external tools) remain; only the terminal and the git interface disappear, so non-developer colleagues can use it comfortably too. You can switch on the Settings → General tab. Source: Google Antigravity (X).
OpenAI: multiple local computers can now be connected to one dot. You can choose per task which machine the agent works on, so for example it can build and test the app on a separate machine while you keep coding on the laptop. Source: OpenAI Developers (X).
Nous Research: Hermes cloud agents are available from the browser, with the same interface as the desktop app. A personal agent started by name and model in the Nous Portal is hosted by Nous in its own container, around the clock, so it can also carry out scheduled tasks. The Plus, Super and Ultra plans include one cloud agent. Source: Nous Research (X).
ElevenLabs: ElevenAgents Campaign Manager for outbound campaigns. The system is given a contact list and a goal — for example a qualified lead or a confirmed appointment — then stays in touch with each contact over time, keeping the history of earlier interactions, until the goal is met. Source: ElevenLabs (X).
Lovable: Jira and Mailjet connectors, and an acquired team for agent features. Finished apps can now also be connected to Jira and Mailjet, Sentry error monitoring already works inside the app, and the Resend email service can be linked with a simple sign-in. Lovable has also bought the team behind the Radish AI agent infrastructure platform; its founders will strengthen development in London and Stockholm. Source: Lovable changelog, Lovable.
Runway: Sightline, scenes that can be composed from any angle in a 360-degree panorama. It builds a 360-degree panorama of any location, from which setups can be cropped at any angle so the surroundings stay consistent in every frame; the panorama can also be exported as a 3D environment. It is available in App Studio. Source: Runway (X).
Smaller launches: free Inkling 2 Small, Spacefast from Automattic, TK from Medium, AI filters on Instagram, a paid AI allowance at Apple. Thinking Machines' Inkling 2 Small model is free for a limited time on OpenRouter and OpenCode, and according to Arena the open weights are coming soon; OpenRouter's MCP server also got a major update. Automattic, the company behind WordPress.com, is launching Spacefast, which gives vibe-coding apps (apps written with AI, mostly from natural-language instructions) a database, sign-in, file uploads, email and analytics; in Medium's writing app TK, AI can be invoked in the editor with “@TK”. With Instagram's Effect Builder, US adult users can make a Stories filter from a text description, and Apple has introduced a $49.99 Intelligence Usage Pack in Logic Pro to expand the quotas for cloud AI features, with the other Creator Studio apps following this month. Source: Thinking Machines (X), OpenRouter (X), TechCrunch, TechCrunch, The Verge, The Verge.
Policy & safety
OpenAI: Control Room, a central “kill switch” that can stop any research model run within 35 seconds. The system records every model run on the research infrastructure — reinforcement learning, evaluation, agent sampling, data processing — and works with monitors that watch for misaligned behavior: if a monitor or a researcher asks for a stop, the inference engines give that task no further model output. Every agent research run needs an approved safety justification, and the system “fails closed”: if Control Room goes down or is switched off, every request is rejected. It is already live for frontier runs, and will be rolled out to the other research runs in the coming days. Source: OpenAI.
Nvidia spy case: a 4% drop, the first investor lawsuit, and Trump says the FBI is also investigating. Nvidia shares fell more than 4% on Monday, several law firms announced investigations, and the first investor class action was filed in the Northern District of California. According to Trump the FBI is “looking into” Jensen Huang's affair, and Huang told him he had no idea the woman was a spy. According to The New York Times, a Commerce Department investigation of Megaspeed over suspected chip smuggling was moved to the Justice Department after Nvidia intervened; China denies any involvement, and Democratic leaders are pushing for congressional hearings. Source: TechCrunch, AP, The New York Times, CNBC.
Fireworks: hackers accessed customer datasets, unfinished model weights and billing data. According to the company, which specializes in running and customizing open models and is valued at $17.5 billion, “LLMjacking” (reselling stolen model access) took place over the past month, a continuation of an October incident, and on October 24 machines storing research and development data were also breached. Two of Forbes's sources among its customers say they have already largely taken their work off the platform; Palo Alto Networks' Unit 42 is helping with the investigation. Source: Forbes.
Google Gemini: some connected Gmail and Drive content can also be seen by human reviewers. According to a Google help page, a subset of Connected Apps data — excerpts and summaries from relevant emails and files — is reviewed by trained human reviewers for debugging and improvement, and it asks that anyone who does not want a reviewer to see confidential data not connect these apps to Gemini. A GitHub write-up that reached the Hacker News front page says this goes against a 2023 promise, and that personal content can also be fetched in the background without a separate request; the latter is based on user tests, and Google has not commented. Source: Google, GitHub.
Trump on 60 Minutes: a big AI company's CEO was threatened with prison. The president said the risk is “nowhere near as real” as AI companies make it sound, and is manageable; according to him, someone was reported to be “doing something bad” with AI, so they called him and told him to “get ready for jail,” after which he “changed everything.” He did not give a name, only that it was the head of an AI company, “one of the big ones.” He added that he would “absolutely” work with Democrats on AI safety, and does not mind guardrails, but does not want to put the US “deeply in second place.” Source: The Verge.
EU: the Digital Fairness Act draft also targets AI systems that simulate a relationship. According to a draft seen by Bloomberg, companies would have to let adult users turn off addictive features, and the rules would also cover AI chatbots that imitate a friendship or a romantic relationship. Source: Bloomberg.
Palantir: lobbied a UK minister on WhatsApp so its AI platform could also reach GP records. According to messages obtained by Foxglove under a freedom-of-information request, the company's UK lead urged then health minister Zubir Ahmed to also bring GP records, which were expressly excluded from the £330 million NHS contract, into the system, and wrote that it was worth getting ahead of the issue or opponents would exploit it. Organizations representing doctors and patients, including the British Medical Association, oppose linking GP records, saying it could undermine patient trust. Source: The Guardian.
Meta: AI estimates Instagram users' ages worldwide, to put more teens into protected accounts. The system also puts into a teen account anyone it thinks is young, even if they entered an adult date of birth; it draws inferences from profile text (for example birthday posts and references to a school year) and from photos and videos. The rollout is underway on Facebook, and is coming to Threads later this year. Source: TechCrunch.
Business & society
Anthropic: the waiting period is over, and the IPO campaign could start as soon as Tuesday. According to Bloomberg, the company could begin formally approaching investors (the roadshow) on Tuesday, even though AI stocks wobbled last week. An IPO analyst said last week's market swing pushed the volatility index into a range in which companies usually postpone going public. Source: Bloomberg.
Google: TPU 8tx and TPU 8i Max, four TPU systems in a single generation, in one year. TPU 8tx adds confidential computing: data and model weights stay protected even while in use, and the chip works in a pair with an AMD processor. TPU 8i Max puts two 8i chips in one package, with twice the compute, memory and memory bandwidth for long-context work. These are not a new chip generation, but new system configurations built on the same silicon. Google counts nearly 100 external TPU customers, and has open-sourced TorchTPU so PyTorch code can run on TPUs with few changes. Source: Google, The Decoder.
OpenAI: ChatGPT Enterprise can now also be bought on the AWS Marketplace. Companies can buy it through their existing AWS purchasing and billing relationship, and can count the spend toward their AWS commitment; the package includes ChatGPT, ChatGPT Work, Codex and dots, with the same features as buying directly from OpenAI. ChatGPT for Healthcare and ChatGPT for Financial Services are also available this way. Source: OpenAI.
SpaceX: a gas plant and supercomputer hub at Starbase, about $24 billion for the first phases. Bloomberg cites a public hearing notice for the plan; according to The Information, because of delays in Texas power permits, Musk's companies may already be paying for leased chips before the host facilities are finished. Source: Bloomberg, The Information (X).
Disney and Adobe: their own “artist-led” generative models in Firefly Foundry. Under the multi-year agreement, Adobe will build models tailored to Disney characters, environments and visual styles, safe for commercial use, for the studio's production workflows, from theme parks to ESPN content. Source: The Verge.
Nvidia: investment in Korean AI startups in Seoul. At a forum with South Korea's science ministry, Nvidia announced an investment in Point2, which develops radio-frequency chip interconnects, and plans further investments in Motif Technologies, RLWRLD and Upstage. Nvidia also released dedicated Dynamo deployment recipes for running leading Korean open models in production. Source: Nvidia, Nvidia.
Fundraisings: Nominal reached a $3 billion valuation, Mavria $1 billion. Nominal, which is building an AI platform for hardware testing, tripled its valuation in eight months with a $240 million Series C; Mavria, which designs modular data centers that can be delivered by truck, raised $130 million led by Paradigm, and Israeli AI publishing company Spines raised $29 million in equity and $35 million in debt. Source: TechCrunch, WSJ, Globes.
New metrics: the second version of the Artificial Analysis Decision Index and Epoch's inference performance index. Artificial Analysis's index ranking decision models is now 45% based on private datasets, and also rewards a model that is well calibrated, meaning its confidence matches its accuracy; Drex Max v1.1 and GLiDE v1.2 lead with 73 points. Epoch's new metric also counts memory bandwidth and inter-chip communication, and according to it raw compute underestimates the GB300's inference advantage over the H100 by a factor of ten. Source: Artificial Analysis (X), Epoch AI (X).
Rumors
Google: tests of a new Gemini variant called “Krypton” are circulating. According to one tester, Krypton thinks about twice as long on the same task as earlier models, and from a single instruction, with no external assets, built an animation scene in pure code; the demos were also picked up in testingcatalog's daily roundup. Google has not confirmed it. Reliability: low–medium (independent user tests, no official confirmation). Source: chetaslua (X), testingcatalog (X).
Anthropic: Fable 5.2 is reportedly the oldest of the recently tested checkpoints. According to legit_api, the company is also working on newer Fable, Opus and Sonnet variants; there are no specifics. Reliability: low–medium (a single code-tracking source that has been accurate several times before). Source: legit_api (X).
OpenAI: a hidden 1-million-token model code-named “Wyvern” was spotted in Devin. Citing a user, testingcatalog writes that a model routed through OpenAI appeared in Cognition's agent. Reliability: low (second-hand detection). Source: testingcatalog (X).
Was Opus 5.5 weakened? Users say it has been thinking differently since November 3. An X post that also became one of the most-commented AI topics on the Hacker News front page claims that Opus 5.5's thinking patterns changed on November 3; Anthropic has not confirmed this. Reliability: low (user observation, no official confirmation). Source: Shane Melaugh (X).
What this means for you: on Monday a large model and a small one arrived at once, and both got cheaper: Fable 5.2 is built for hard, open-ended tasks, Muse Ember for the many repeating agent subtasks. If you are planning to have agents do work at your company, on our AI agents page, you can see which tasks are worth it, and with our vendor-independent consulting, you can choose a model and a provider so that a price change or a better model does not tie your hands.