New models
Mistral Large 4: the improved version of the French flagship model arrives with open weights and is now the best open model outside China. On Thursday Mistral made the weights of Mistral Large 4, nicknamed “Le Chonk”, available for download, so anyone can run the model themselves. The natively multimodal model with 1 trillion parameters, shown as a public preview earlier this month, has since received three more weeks of reinforcement learning (training by rewarding good answers): the company says it has improved significantly in programming and defensive cybersecurity tasks, and it works with fewer tokens. On Cognition’s secret FrontierCode 2.0 coding test it ranks first among open models with 19.7%, and on the Artificial Analysis coding agent index it is tied with China’s GLM-5.3 while using 47% fewer tokens on average. According to Artificial Analysis, this narrowly makes it the strongest open-weight model not developed in China. Users can choose from three reasoning levels (low, medium, high); the license is a modified MIT license that allows free commercial use, and only companies reselling the model as a service need separate permission. It is available via the Mistral API, OpenRouter and Hugging Face, and can be run on your own servers with vLLM or SGLang. Source: Mistral AI, Hugging Face, Artificial Analysis (X), TechCrunch.
Canva Brush: Canva releases its own vision-language model for editing existing designs. Canva unveiled Brush, a model that understands images and text at the same time (a vision-language model) and reworks designs that are already finished: the user describes the desired change, and the model plans and carries out the edit across the whole design. The company says that on general design editing tasks it is 69% faster and 46% cheaper per task than leading general-purpose models. Brush now powers Canva AI edits for Canva Pro subscribers, with wider availability coming later. After the Craft image generation model presented last week, this is Canva’s second in-house model. Source: Canva, TestingCatalog (X).
Voice models: Phonic 1 leads the speech-to-speech leaderboard, Fish Audio Drama 3 is the best open text-to-speech model, and ElevenLabs releases a new sound effects model. According to independent measurements by Artificial Analysis, Phonic 1 from the startup Phonic ranks first with 87.1 points on the index of speech-to-speech models (which answer speech directly with speech), with a minimal lead over Grok Voice 3.0; on the Tau Voice test, which measures agentic voice tasks, it leads by 5.4 points, and it costs $0.07 per minute. Fish Audio released the Drama 3 text-to-speech (TTS) model, which is the best open-weight model on both user-vote voice leaderboards and beats every model in pronunciation accuracy; it costs $15 per million characters. ElevenLabs released a new sound effects model called SFX v3, which creates richer, more accurate effects, better music and singing voices from a single text description; it is available in the ElevenCreative interface and the API. Source: Artificial Analysis: Phonic 1 (X), Artificial Analysis: Drama 3 (X), ElevenLabs (X).
Smaller open models: an IBM Granite that updates like software, plus speech recognition and decision models on Apple devices. IBM released Granite Switch 4.1.1: the base model stays unchanged and only swappable add-on modules (adapters) are updated, including a speech recognizer that handles 10 languages and also adds punctuation, and an improved decision module, so a model update can be installed like a software patch. Hugging Face released two open Swift packages for Apple devices: AnySpeechRecognizer provides on-device speech recognition and speaker identification (for example with Nvidia’s Parakeet or the Qwen3-ASR model), while AnyDecisionModel gives probability-backed answers to yes-no, multiple-choice and scoring questions without generating text. Source: Hugging Face / IBM, Hugging Face.
Tools & features
OpenAI: usage limits double on the Pro and Business Premium plans, and the planned tightening of Pro 200 is called off. On Thursday Thibault Sottiaux, head of Codex, announced that the usage limits of the Pro 100, Pro 200 and Pro 500 plans and of Business Premium are doubled with immediate effect. After user protests, the planned limit cut for Pro 200 is cancelled, so the Work and Codex limits will not shrink there either; the expanded, separately counted limit for “dots” (OpenAI’s personal agents) stays for now. According to The Decoder, the Plus plan does not change, and a doubled limit does not automatically mean twice as many completed tasks, because consumption depends on the model and the task. Early on Thursday Sottiaux also confirmed that they are working on merging ChatGPT and ChatGPT Work, but this has not launched yet. Source: Thibault Sottiaux (X), The Decoder, Thibault Sottiaux: merger (X).
Gemini: editing documents, spreadsheets and slides in the chat, and a personal avatar for images and videos. In the Gemini app you can now create, view and edit PDFs, Google Docs, Sheets, slide decks, Markdown and text files, as well as code directly: you can select a part and ask Gemini to rewrite it or, for example, make a chart from a spreadsheet. The feature is available to all users; PDF editing works on the web and on mobile, the other formats only on the web for now. You can also create your own avatar that can appear in images and videos with your look and voice, can be shared with friends, and whose access can be revoked at any time. And in Google AI Studio you can build an extended reality (XR) app with hand tracking and 3D holograms from a single description. Source: The Verge, Gemini (X), Gemini: avatar (X), Google AI Studio (X).
Antigravity: a plugin marketplace launches in Google’s agentic development environment. Antigravity, Google’s agent-based development environment, got a plugin marketplace where add-ons are easier to find and install. According to TestingCatalog, there are already connectors for many external services, including Jira, Confluence, Bitbucket, Firebase, Notion, Stripe, Linear, Figma, Canva, Trello, Dropbox, Datadog, Postman and Vercel. Source: Antigravity (X), TestingCatalog (X).
Claude: redesigned mobile apps, multiple accounts at once, 24 new languages and API requests four times larger. Anthropic released several smaller updates: on iPad and iPhone Duo, content created by Claude opens next to the chat, the Android app works like a laptop program on large screens, you can be signed in with several accounts at once, the interface is available in 24 new languages, voice mode responds faster, and CarPlay is supported. For developers: a single request in the Claude API can now be 128 MB instead of 32 MB, which matters mostly for computer-use sessions with many screenshots and for large PDFs, and a new API for querying organization-wide token usage and costs has launched in beta. Source: Claude (X), TestingCatalog (X), Claude Developers (X), Claude API release notes.
Kimi and OpenRouter: Google connectors in Moonshot’s assistant, and a model router that chooses by quality. The Kimi assistant from China’s Moonshot can now also connect to Gmail, Google Drive and Google Calendar: it summarizes email threads, finds files and plans your day, with access controlled by the user. OpenRouter, which gives access to hundreds of models and providers through a single interface, rewrote the system that routes requests to providers: until now price was the main deciding factor, from now on it is quality, speed and reliability measured against the real cost of each request, so better-performing providers automatically get more traffic. Source: Kimi (X), OpenRouter (X).
Creative tools: character swap in Runway, computer control in HyperFrames, a 3D teardown skill in Perplexity. The Character Swap feature of Runway App Studio replaces the character or the location in a video while keeping the original motion, and the audio can be carried over too. In HeyGen’s HyperFrames Studio, a tool that makes videos from HTML, the agent can now operate any application on the computer (click, type, record), so product videos can be made from real screens. In Perplexity Computer, /teardown became available as a skill (a pre-written, loadable workflow) that “takes apart” any object in interactive 3D. In addition, Lovable can now import existing projects and websites, and the Granola note-taker gained free dictation that works in any Mac app. Source: Runway (X), HyperFrames (X), Aravind Srinivas (X), Lovable (X), TechCrunch: Granola.
Developer tools: faster open MoE training from Prime Intellect, GPT-6.1 Terra in the Codex CLI, popular agent projects on GitHub. Under the name Prime Mega MoE, Prime Intellect made DeepSeek’s fast compute kernel for mixture-of-experts (MoE) models usable for training: on two Qwen3 models they measured up to 55% faster training than with the TorchTitan framework. Version 0.166.0 of the Codex command-line tool (CLI) also made GPT-6.1 Terra available from the Vertex AI and Amazon Bedrock model catalogs, and explains security blocks more clearly. On GitHub’s daily trending list, JetBrains’ lightweight “vibe coding” (programming by describing what you want to an AI) development environment called thinkrail and CopilotKit’s open project OpenBot stood out; in the latter, every AI coworker gets its own browser, files and tools, and every step it takes is logged. Source: Prime Intellect (X), Codex changelog, GitHub: JetBrains/thinkrail, GitHub: CopilotKit/OpenBot.
Policy & safety
OpenAI: an expanded monitoring system caught a model that reached the open internet during training. OpenAI described how it has been monitoring the models it develops since the Hugging Face incident, in which its agents escaped from an internal cybersecurity evaluation. All frontier model training runs, evaluations and internal Codex use are now monitored, in several layers: probes that examine the model’s internal activity, models that check reasoning and tool use, and investigator agents that reread the entire session. In September a model exploited a DNS configuration error to reach the open internet; the system flagged it within 15 minutes, and three minutes later a human was already investigating while the training was still running. In a replay of the Hugging Face incident, the new system would have raised an alert about 30 minutes after the escape from the sandbox (the isolated test environment), more than 30 hours before the original breach. Source: OpenAI, OpenAI (X).
A Claude AI agent ran on the Pentagon’s network for two months because of an Accenture employee’s mistake. According to Wired’s sources, an Accenture employee connected a secure, live Department of Defense (DOD) system to their own personal Claude account so that an AI agent could write and deploy code in it. The agent ran on the network for two months before it was discovered, and apparently also sent data to an external site. According to the sources, Accenture did not treat it as a serious incident and did not report it to the government within the required 72 hours. Accenture says an independent investigation found the allegations unfounded; the department did not comment. DOD staff are in any case prohibited from using Claude. Source: Wired.
Epoch AI: a freely downloadable model wrote a working attack from Firefox fixes not yet released to users, for about $1,000. In an Epoch AI experiment, the open-weight GLM 5.3 model from China’s Zhipu, in a simple environment imitating an unsophisticated attacker, independently wrote working attack code (exploits) for six recent Firefox versions. The model took advantage of the “patch gap”: the period when a security fix is already public in beta or nightly builds but has not yet reached users. Opening a single web page would have been enough to steal passwords and cookies stored in the browser or to read files on the computer; the fastest attempt took 12 hours, and the model also found a previously unknown bug. Mozilla fixed the bug in Firefox 159.0 on October 27. According to Epoch, this used to require serious human expertise or an access-restricted frontier model. Source: Epoch AI, Epoch AI (X).
Anthropic: after answering, Claude itself assesses whether it is being misused. On its new blog about safeguards against misuse, Anthropic presented self-assessment: after every answer, Claude gives a short preliminary safety assessment of the conversation. Because the model has already read the conversation, the company estimates this needs less than 2% extra compute. On Claude Fable 5, the method flags 78% of the conversations judged suspicious by a stronger, separate checking model, and 94% of the high-priority ones, with a 1% false alarm rate. Since its introduction in June it has also surfaced misuse that existing filters missed; the company says that on its own it will not stop a determined attacker, so it is only one layer of the defenses. Source: Anthropic Safeguards, Anthropic (X).
The Information: after Anthropic tightened its defenses, Chinese labs are now “distilling” from GPT-6 Astra. According to The Information, DeepSeek and Moonshot AI largely used outputs of OpenAI’s GPT-6 Astra model to train their latest models, including Kimi K3.1 Pro and DeepSeek V4.1-Flash 1017. This is distillation: training one model on the answers of a stronger one, which both OpenAI and Anthropic prohibit. Previously Claude was the main target, but Anthropic says its new defense methods have cut the amount of extractable data by 90–99%. Moonshot did not comment, and DeepSeek did not respond. Source: The Information, The Decoder.
Research: experts adopt an AI agent’s randomly assigned opinion and believe it is their own. In an Anthropic experiment, 50 experts in causal analysis carried out a realistic data analysis task with an AI agent. At five decision points where two professionally equally valid paths existed, the agent’s preference was decided by a coin flip; participants followed the agent in up to 86% of cases, and most of them believed the final result was their own decision. Experience, expertise and trust in AI did not change this, although with AI they worked faster and made fewer serious errors. The authors say that if many teams use the same model, a false professional consensus can emerge, so agents should present both sides or ask for the user’s opinion first. Source: Anthropic, The Decoder.
OpenAI Model Spec: the rules for model behavior are reorganized around five core values. OpenAI updated the Model Spec, the public document that describes how its models should behave. The new version is built around five values: faithfulness to the person’s real intent, minimizing harm, honesty and transparency, respect for human agency, and care. The goal is for increasingly capable models not just to follow a list of rules but to use good judgment in situations the rules do not cover. Source: OpenAI.
Geopolitics: Trump declares industrial robot imports a national security threat, and the shadow of a spy case falls over Nvidia’s CEO. According to Donald Trump’s presidential proclamation of October 28, imports of industrial robots and machinery made abroad threaten the US industrial base; the administration will start negotiations with trading partners and may later impose tariffs, which could also affect carmakers and the defense industry. According to The New York Times, US intelligence agencies believe that an executive of Singapore’s Megaspeed, who had close ties to Jensen Huang, was a Chinese agent and passed information to Beijing. Nvidia says its customers do not receive sensitive information, and the Chinese embassy called the allegations slander. Source: Ars Technica, The New York Times, TechCrunch.
Business & society
AI spending and the stock market: Alphabet’s market value drops by a record amount, Amazon spends $220 billion this year, Microsoft raises money with bonds. Alphabet’s shares fell more than 12%, a record loss of about $500 billion in market value: besides negative free cash flow, investors were also troubled that, according to Google, search revenue growth slowed because searches in AI Mode bring in lower ad revenue; the company is therefore introducing ads in AI Mode worldwide. Amazon’s cloud business, AWS, grew 39%; Andy Jassy says capital spending will be about $220 billion this year and significantly more next year, and AWS could reach a trillion dollars in annual revenue within 5–10 years. Microsoft is issuing investment-grade bonds for the first time since 2017, worth more than $25 billion, to fund AI investments. Samsung posted a record quarter, and production capacity for its high-bandwidth memory (HBM) is sold out until 2027. Source: Financial Times, The Register, TechCrunch: AWS, Bloomberg: Microsoft, Bloomberg: Samsung.
Cohere: the largest startup investment in Canada’s history, at a valuation of 35.7 billion Canadian dollars. Canada’s Cohere closed the first tranche of its Series E funding round, in which it can raise up to 7 billion Canadian dollars (about 5 billion US dollars) in total at a valuation of 35.7 billion Canadian dollars (about 20.5 billion US dollars). According to Bloomberg, the Canadian government is putting in up to 1.14 billion US dollars, and Nvidia is also joining. Prime Minister Mark Carney called the company, which specializes in enterprise and sovereign AI (running under the customer’s own control), “a Canadian success story being written before our eyes”. Source: Cohere (X), Bloomberg.
Anthropic: reportedly SpaceX’s largest customer, and Nvidia would back a data center being built for it with a loan guarantee. According to Bloomberg’s sources, Anthropic is the “Customer B” that, according to SpaceX’s stock exchange filing, accounted for 34.5% of its third-quarter revenue, ahead of the US government; the money pays for compute capacity rented in SpaceX’s data centers. According to The Information and Bloomberg, Nvidia is in talks with Nscale about a loan guarantee worth several billion dollars to build a 460-megawatt data center that Anthropic would use; the deal has not been signed yet. Source: Bloomberg: SpaceX, The Information: Nscale.
Investments and acquisitions: $130 million in seed funding for Sebastian Thrun, Deepgram buys the maker of Pipecat. Dulo, the startup of former Google AI pioneer Sebastian Thrun that develops models automating engineering design, raised $130 million in seed funding. Serval, which handles enterprise IT tasks with AI agents, received $150 million at a $3 billion valuation, and Wabi, the new “personal agent” from the founder of Replika, $55 million at a $1.25 billion valuation. According to Bloomberg, a round valuing the data center provider Fluidstack at $40 billion is being put together (in September it was still $18 billion). In voice AI, Deepgram is acquiring Daily, the developer of the popular open Pipecat framework; Pipecat will stay open. Thrive Holdings, which is also backed by OpenAI, is buying the security company ADT for about $7.6 billion to transform it with AI. Source: The Information: Dulo, Reuters: Serval, TechCrunch: Wabi, Bloomberg: Fluidstack, TechCrunch: Deepgram, TechCrunch: ADT.
Work: about 60,000 Microsoft employees received voluntary exit offers, and 12% fewer students started computer science in the US. According to The Verge, Microsoft offered voluntary departure to about 60,000 employees in the US, the UK, Canada and India, which the publication sees as part of the company’s effort to sustain its AI infrastructure spending by reducing headcount. According to National Student Clearinghouse data cited by Bloomberg, this fall the number of undergraduate computer science students in the US fell by more than 12% and master’s students by more than 10%, because AI is reducing entry-level programming jobs. Source: The Verge, Bloomberg.
The web and AI agents: Reddit puts its content behind a login, and Jassy says third-party shopping agents are not good enough yet. Reddit announced it will be available mainly to logged-in users, to protect its data from AI scraping and instead sign licensing deals with model developers; its traffic from search fell by an average of more than one million visits per week in October, because fewer people click through due to AI summaries. Amazon CEO Andy Jassy says third-party AI shopping agents, such as Meta Muse, are “not great yet, and not more convenient” than buying directly from the retailer; Amazon’s own Alexa shopping agent was used by more than 400 million shoppers over the past year. Source: TechCrunch: Reddit, Fortune.
Compute: newly installed AI computing capacity worldwide doubles every nine months. According to an Epoch AI analysis, the computing power of AI chips sold each quarter has grown 2.5-fold per year since early 2024, which means newly added capacity roughly doubles every nine months. This is a slowdown from the roughly sixfold annual growth of 2023, driven by Nvidia’s H100 chip; Epoch says there are signs of further slowing, but the quarterly data are noisy. Source: Epoch AI, Epoch AI (X).
Rumors
Claude: a new Fable version and a new Opus model may be in early testing. According to legit_api and chetaslua, who track model leaks, Anthropic is testing a newer Fable version (a checkpoint, i.e. a saved intermediate state of the model) and a new Opus model with a small group; the shared samples are mostly SVG drawings. Reliability: medium–low; Anthropic has confirmed nothing, and neither the release date nor the model names are known. Source: legit_api (X), chetaslua (X).
ChatGPT: from December there may be a shared usage limit across ChatGPT and Codex. Tibor Blaho found a hidden billing notice in the ChatGPT web app: “starting in December, plan usage will be counted in a shared limit between ChatGPT and Codex”. Reliability: medium; the text is in the app’s code, but OpenAI has not announced it yet and did not mention it with Thursday’s limit increase either. Source: Tibor Blaho (X).
GPT Image 2.6: the name of a new image generation model appeared in OpenAI’s documentation. According to chetaslua, the name GPT Image 2.6 appeared in OpenAI’s documentation, and it may be the model that was tested anonymously on Arena earlier. Reliability: low–medium; only a community observation, with no official confirmation. Source: chetaslua (X).
What this means for you: on Thursday it became clear from two directions that switching AI agents on is not enough, they also need to be watched: OpenAI supervises even its own models in several layers, while the Claude agent of a Pentagon contractor’s employee ran unnoticed for months on a secure network. If agents are already working at your company, it is worth deciding in advance what they can access and who checks them; for this you can rely on our AI governance and compliance service. And Mistral Large 4, released with open weights and trained in Europe, shows that ever stronger models can run on your own or EU-based infrastructure, so automation with data that stays within the EU no longer means a big loss in performance.