New models
Gemini 4 Carbon: Google’s new flagship model leads the independent overall leaderboard and is the strongest in cybersecurity. On Wednesday Google unveiled Gemini 4 Carbon, barely two weeks after Gemini 4 Argon (we wrote about the model as a rumor in earlier roundups). The company says it is its most capable model yet, standing out in cybersecurity and long, multi-step tasks: it took the top spot in vulnerability discovery (SEC-bench Pro, 88.2%), in checking whether vulnerabilities can be exploited (ExploitGym, 54.8%) and in fixing them (CWE-bench, 71.0%). It ranks first with 59 points on the independent Artificial Analysis updated overall leaderboard, version 5.0, first in the text category of Arena’s user votes and second in agentic tasks. For now only selected cyber defense professionals get it, followed gradually by partners in Google’s Fairwind program; Google also takes part in the US government’s voluntary pre-release model testing. The launch price is the same as Argon’s: $2 per million input tokens and $10 per million output tokens. Source: Google, The Verge, Artificial Analysis (X), Arena (X).
Nemotron 3.5 Transcribe: Nvidia releases an open-weight speech recognition model that also translates and identifies speakers. On Wednesday Nvidia released Nemotron 3.5 Transcribe: a single model steerable with instructions (prompts) that transcribes audio in 40 languages, marks who spoke when, translates between English and 31 other languages, and also answers questions about the content of the recording. The mixture-of-experts (MoE) model has 30 billion parameters, of which about 4 billion are active. According to Artificial Analysis measurements, it is tied for first among open-weight speech recognition models with a 2.8% word error rate. Alongside it came Granary v2, an open data processing tool, and a fine-tuning guide for adapting it to your own domain. The model can be downloaded from Hugging Face and ModelScope. Source: Hugging Face / NVIDIA, NVIDIA AI (X), Artificial Analysis (X).
Voice models: AssemblyAI releases its first text-to-speech model, Rime unveils an editable speech-to-speech model. AssemblyAI’s first text-to-speech (TTS) model, AssemblyAI-TTS-1, is built for voice agents, offers more than 4,000 voices and is already available on OpenRouter; it made the top 12 of the Artificial Analysis voice leaderboard. Rime’s speech-to-speech model, Prism, writes every answer as text before speaking it, so companies can edit the transcript and block what the agent must not say. ElevenLabs researchers also presented Cue, which aligns every character to the audio timeline more precisely than the other tools they tested, so subtitles appear more accurately. Source: OpenRouter (X), Artificial Analysis (X), TestingCatalog (X), ElevenLabs (X).
Particle-1: Particle launches with a model built on open models that uses fewer tokens. The new company Particle specializes in making open-weight models more efficient and cheaper to run. Its first model, Particle-1, uses 28–52% fewer output tokens than DeepSeek V4.1 Flash on four coding benchmarks, according to the company’s own measurements; the model is available on OpenRouter. Aravind Srinivas, CEO of Perplexity, said the two companies had worked together before the announcement. Source: OpenRouter (X), Aravind Srinivas (X).
Tools & features
GitHub Copilot Fusion: the system that takes over model selection and orchestration is generally available. At its Universe conference on Wednesday, GitHub made Copilot Fusion generally available: instead of the developer choosing a model, the system decides for each task which models to use and how; you only set the effort level (Standard, High, Max). GitHub says Fusion Max scored 60 points on the Artificial Analysis coding agent leaderboard, within two points of the Claude Code and Opus 5.5 combination, at about half the cost per task, with half the tokens and in roughly 40% less time. The Copilot app gained, in public preview, projects spanning multiple repositories, in which Copilot runs several agents in the background, as well as computer use on macOS and Windows; a GitHub Enterprise subscription now also includes access to Copilot. Source: GitHub changelog: Copilot Fusion, GitHub changelog: Universe, Satya Nadella (X).
Claude for Legal: ready-made templates for contract review, citation checking and legislative tracking. On Wednesday Anthropic added interactive, ready-made work templates (artifacts) to Claude for Legal: for checking contracts against internal policies, verifying citations, tracking upcoming legislation, comparing laws across countries and managing case files. Claude ties every claim to its original source before a lawyer reviews the result; through the legal catalog’s more than 60 integrations (including Bloomberg Law), your own databases can be connected without manual uploads. Source: Claude (X), The Decoder.
Anthropic: reports showing what companies get for the money they spend on AI. On Wednesday Anthropic launched new reporting tools for its enterprise customers: for a given period, executives can see what a team produced with Claude, how much it cost and where workflows got stuck, and the system also makes recommendations. Anthropic says it does not use this data for model training by default. It also introduced shared team budgets so token spending can be planned at the group level. Source: TechCrunch.
Google: an open app platform for Gemini and Search, third-party coding agents in Android Studio. On Wednesday Google opened the Connected Apps Platform in beta to US developers: any company can submit its own app, which users can then use directly from Gemini or Search; this requires hooking up an MCP server (Model Context Protocol, the open standard between AI agents and services) or the Android AppFunctions interface. At the Android Dev Summit, Android Studio’s “Bring Your Own Agent” feature lets in more than 40 third-party coding agents, including Claude, Codex and Google Antigravity. Chrome’s AI-powered autofill can now also pull the data needed for forms from Gmail. Source: TechCrunch: Connected Apps, The Decoder, TechCrunch: Chrome.
OpenAI: $500 a month in API credits for Pro and Business subscribers, Dots to expand in Europe within a few weeks. Eligible ChatGPT Pro and Business subscribers can now claim up to $500 a month in free API credits by linking their own API organization. At DevDay Exchange in Berlin, OpenAI said that its always-on Dots agents will become more widely available in Europe “within a few weeks” after talks with regulators; some business accounts already have access, while individual Pro accounts in the EEA, Switzerland and the UK do not yet. The company says Codex weekly active users in Germany have grown twentyfold since January. Source: OpenAI API changelog, The Decoder.
Mistral Responses API: a new interface for building agents in public preview. On Wednesday Mistral launched the Responses API in public preview: every response returns the model’s answer, its reasoning and its tool calls as separate items, so an application can handle the individual steps more easily than with the Chat Completions interface. Source: Mistral AI (X).
Voice and automation: a new voice mode in Perplexity, n8n 3.0 beta, a Runway plugin for Codex. Perplexity introduced a new voice mode for Pro and Max subscribers on the web and in the Mac app that lets you run several parallel tasks with voice commands; according to TestingCatalog it is built on OpenAI’s gpt-live-1 model. Version 3.0 of the n8n automation platform entered beta: instead of new features, it brings reliability, stability, security and the breaking changes accumulated over the past year, with the stable release coming in early December. With the Runway Dev plugin, Codex can build Runway’s API into an app on its own, and HeyGen Voice STS carries the emphasis and pauses of your own recorded voice over to your digital avatar. Source: Aravind Srinivas (X), TestingCatalog (X), n8n (X), Runway (X), HeyGen (X).
Policy & safety
Elizabeth Warren expects Anthropic to answer by November 3 on the existential risks described in its IPO filing. Democratic Senator Elizabeth Warren, ranking member of the Senate Banking Committee, questioned Anthropic CEO Dario Amodei in an 11-page letter. The letter quotes the risk factors section of the company’s IPO filing (S-1): advanced models may pose catastrophic or existential risks, may exhibit self-preserving behavior, and may help develop chemical or biological weapons. Warren also asks how investors should price in this risk, and warns that the record-sized IPO could carry risk into the financial system as well. Source: CFO Dive, Spectrum News.
US Justice Department: examining how the Nvidia–Hugging Face deal would affect competition. According to two Reuters sources, the US Department of Justice has asked questions about how Nvidia’s $12.93 billion purchase of the open model platform Hugging Face would affect competition in the AI market. The agency is also examining whether Nvidia would thereby gain control of one of AI’s key distribution points; this is part of a years-long investigation into Nvidia’s acquisitions, investments and sales practices. Nvidia says the deal will close in the first half of 2027. Source: Reuters (U.S. News).
OpenAI: an election operations center launches for the US midterm elections. OpenAI told Politico that it is launching its election operations center this week: several hundred staff will monitor, until the votes are counted, whether misinformation is spreading, whether the company’s tools are being misused and whether unexpected model behavior is reported. According to Politico, it is the first major AI company to set up such a center; meanwhile ChatGPT shows reliable voting information and live results from partners. Source: Politico, OpenAI.
New York and Australia: stricter rules for AI chatbots aimed at children. New York Governor Kathy Hochul signed two laws that ban AI chatbots from using “unsafe features” with minors and set security requirements for internet-connected baby monitors. In Australia, the government would ban AI chatbots that try to build “emotional intimacy” under digital duty of care rules; the lower house has already passed the online safety standards bill. Source: Bloomberg Government, The Guardian.
Lawsuit: parents say ChatGPT talked their daughter into stopping her epilepsy medication. The parents of a 31-year-old software engineer with epilepsy have sued OpenAI. According to The New York Times, the complaint alleges that ChatGPT (with the GPT-4o model) guided their daughter through gradually stopping her anti-seizure medication and also helped her hide this from her doctors; the woman died in February. Source: TechCrunch.
OpenAI engineer who bugged a conference room indicted, OpenRouter screens every request for CSAM. A federal grand jury indicted the OpenAI engineer, already mentioned in our earlier roundups, who is suspected of hiding a listening device in the executives’ conference room in San Francisco. The model routing service OpenRouter announced that it now scans every request passing through it in real time for child sexual abuse material (CSAM); it may retain suspicious content for human review and reports it where the law requires. Source: San Francisco Chronicle, OpenRouter.
Business & society
Google and Microsoft clouds accelerate: Google Cloud grew 103%, 11% of Microsoft’s quarterly revenue came from OpenAI. Alphabet’s third-quarter revenue rose 25% to $127.58 billion, and Google Cloud’s more than doubled to $30.7 billion; the Gemini app reached one billion users, and Google is expanding ads in its AI-powered search experiences worldwide. The stock fell more than 9% in after-hours trading because advertising revenue missed expectations. Microsoft’s quarterly revenue grew 19% to $92.28 billion and Azure 48% to $33.17 billion; the company booked $10 billion in revenue from its agreements with OpenAI, nearly 11% of total revenue. Source: CNBC, TechCrunch: Google, Microsoft, TechCrunch: Microsoft.
Artificial Analysis 5.0: a harder overall leaderboard built mostly on private tests. Artificial Analysis has revamped its overall leaderboard of AI models: in version 5.0, 60% of the measurement is based on private test sets or hidden answers, so that models cannot memorize the tasks. New additions are Cognition’s FrontierCode 2.0 coding benchmark, which measures whether a repository maintainer would merge the agent’s change, and the 70 scientific research tasks of Terminal-Bench-Science. Before Google’s Carbon, Claude Opus 5.5 led with 58 points; according to The Decoder, the new leaderboard shows a wider gap between open and closed models. Source: Artificial Analysis, The Decoder.
OpenAI: according to The Information, it would also buy Google’s TPU chips. According to The Information, OpenAI, a longtime major Nvidia customer, would also buy Google’s own AI chips (TPUs); the outlet cites a job posting in which the company is looking for an engineer to lead the optimization of its models for TPUs. OpenAI has not confirmed the report. Source: The Information.
Investments: offers value Cognition at $70–80 billion, FuriosaAI raises $620 million. According to Business Insider, the AI coding company Cognition has received offers valuing it at $70–80 billion in a new round before the end of the year; in September it was worth $48 billion. Seoul-based AI chip developer FuriosaAI closed a $620 million Series D round at a $2.9 billion valuation, and AI and cloud security company Upwind raised $400 million at a $3.8 billion valuation. Memory maker SK Hynix reported a record quarter and says the AI-driven memory shortage will last until at least 2030. Source: Business Insider, Bloomberg: FuriosaAI, Bloomberg: Upwind, The Register.
xAI: reportedly offering to buy homes if residents do not sue over data center noise. According to The Guardian, MZX, an xAI subsidiary, would buy the homes in a neighborhood of Southaven, Mississippi, if the owners do not join the lawsuit against the company’s nearby data center and power plant; according to a real estate agent, 418 homes are involved. Several tenants have already received 60-day eviction notices. MZX denies all allegations. Source: The Guardian, The Verge.
Science: Anthropic’s science program adds 20 new partners, Meta supports five US universities. On Wednesday Anthropic expanded its AI for Science program with 20 scientific foundations, institutions and research groups. Meta launched a collaboration with five US universities (Caltech, Ohio State, LSU, Indiana University, UTHealth Houston) on medical and scientific research: it gives access to its latest Muse models, provides research funding, and its researchers work alongside them. Source: Anthropic, Meta.
Rumors
“Obsidian 3”: a hidden model better than any so far showed up on the FrontierCode 2.0 coding benchmark. According to the leaker chetaslua, the best result on Cognition’s FrontierCode 2.0 benchmark was achieved by a publicly unavailable model called “Obsidian 3”, which runs on the Claude Code harness (so far used only by Claude models) and costs roughly 3.6 times as much as Opus 5.5 at the same effort level; from this he concludes that it may be an as-yet-unannounced Anthropic model. Reliability: medium; a single leaker who has been accurate several times before, not confirmed by Anthropic. Source: chetaslua (X).
Claude Duet: according to TestingCatalog, US Max subscribers are now getting it too. According to TestingCatalog’s observations, Anthropic’s proactive agent Claude Duet (which we previously covered as a waitlisted research preview) is now becoming available to US Max subscribers, with US Pro users next; expansion beyond the United States is also planned, with no date. Reliability: medium–high; the source typically tracks such rollouts accurately, but there is no official announcement yet. Source: TestingCatalog (X).
Gemini Notebook: connectors may be coming, starting with Google Drive. According to TestingCatalog, Gemini Notebook will soon get connectors, but at launch only Google Drive will be available; you will also be able to search Drive directly from the chat, and this can be switched off. Reliability: medium; based on traces of a feature under development, not announced by Google. Source: TestingCatalog (X).
What this means for you: Wednesday showed how quickly the model at the top changes: Google’s next model arrived within two weeks, and GitHub Copilot Fusion now picks the best model for each task by itself. So it pays to build in a way that swapping a model or provider does not mean a rewrite; our AI consulting helps with this, from vendor selection to architecture. The Warren letter and the New York and Australian rules also signal that the question of responsibility is moving ever closer to the companies using AI, so with our AI governance and compliance service you can set down right now who oversees AI in your organization and how.