New models
Genie 4: Google DeepMind unveils a more realistic world model that also generates sound and can be steered on the fly. On Tuesday Google DeepMind unveiled Genie 4, its most realistic interactive world model yet (world model: an AI that generates an explorable virtual environment from a text description, reacting in real time). The new version also creates sound to match the scene, and the world can be changed on the fly with text instructions. The simulation is supported by Gemini 3.8 Flash, and each experience lasts about two minutes, twice as long as with Genie 3; Google says details can still drift in longer runs. For now it is available as a research preview to selected testers, Google AI Ultra subscribers will be able to try it soon (you can sign up on the Project Genie waitlist), and according to Logan Kilpatrick it is also coming to the API. Source: Google DeepMind, The Verge, Sundar Pichai (X).
LFM3: Liquid AI releases six open-weight small models for text, image, audio and decision-making tasks. On Tuesday Liquid AI released the LFM3 model family: six models of roughly 2 billion parameters for text, images, audio, combined handling of all of these (omni) and decision-making tasks, all built on the same base model (LFM3-1.7B-Base). The new architecture, which combines three types of layers (tri-hybrid), was trained on 22 trillion tokens. The company says the 1.7-billion-parameter text model beats the previous 2.6-billion-parameter LFM2.5 with 35% fewer parameters, fits in about 1 GB with 4-bit compression, handles a 256K-token context, and can be further trained on your own data on a single GPU. The weights can be downloaded from Hugging Face, and the models are also available through the company’s API. Source: Liquid AI, Hugging Face.
MAI-Image-2.7: Microsoft’s new image editing model joins the leaders on independent leaderboards. On Tuesday Microsoft AI made the MAI-Image-2.7-Preview image editing model available. It took second place on the Artificial Analysis image editing leaderboard, behind only OpenAI’s GPT Image 2.5 Sunburst, and came third on Arena’s leaderboard based on user votes, 45 points ahead of its predecessor, version 2.6. On the same day MAI-Image-2.6 and the faster MAI-Image-2.6-Flash became generally available (GA) on the Microsoft Foundry cloud platform, with editing from up to five reference images, while MAI Cyber 1.1 Flash, built for vulnerability hunting, delivers the same performance as its predecessor at 18% lower cost per scan, according to the company’s internal measurement. Source: Artificial Analysis (X), Arena (X), Microsoft Foundry, TestingCatalog (X).
Ling-3.1-flash-1019: Ant Group releases an updated open model. Ant Group’s Ling team open-sourced Ling-3.1-flash-1019, an update to Ling-3.1-flash. The mixture-of-experts (MoE) model has 560 billion parameters in total, of which 25 billion are active per token, and it handles a 1-million-token context. It scored 44 points on the Artificial Analysis Intelligence Index, up from its predecessor’s 41; a base model intended for further training (Ling-3.1-flash-base) was also released alongside it. Source: Ant Ling (X), Hugging Face, Artificial Analysis (X).
Tools & features
ChatGPT: now connects to iCloud mail and calendar. On Tuesday OpenAI made the iCloud Mail and iCloud Calendar plugins available in ChatGPT, so from the chat you can search your emails, have them summarized, and manage your calendar too. The plugins are available on all ChatGPT plans on the web, in the desktop app and on iOS. Source: ChatGPT (X).
Apple: an AI assistant is coming to Keynote, Pages and Numbers, open to outside agents too. On Tuesday Apple announced that Keynote, Pages and Numbers are getting an AI assistant in beta, built on Apple’s own models and OpenAI’s: it designs slides, writes and edits text, and creates tables and charts. The three apps are also getting MCP support (Model Context Protocol: an open standard through which outside AI agents can connect to the app), Logic Pro’s assistant can set up a vocal track, and Pixelmator Pro edits images based on text instructions. The features require a device that supports Apple Intelligence and iOS 27, iPadOS 27 or macOS 27; Creator Studio subscribers get the updates free of charge in November. Source: The Verge.
Oracle: ChatGPT and Codex are coming to Oracle’s enterprise applications. On Tuesday OpenAI and Oracle expanded their partnership: ChatGPT and Codex are coming to Oracle’s business applications, so companies using them can work with them on their own data. Oracle’s cloud (OCI) will also start offering Google’s TPU chips, OpenAI’s own Jalapeño chip and Cerebras hardware. The announcement was made at the Oracle AI World conference, where Sam Altman also took the stage. Source: OpenAI, CNBC.
Claude for Financial Advisors: connectors and ready-made workflows for financial advisors. On Tuesday Anthropic released the Claude for Financial Advisors package. The connectors link Claude to the custodians, client relationship management (CRM), portfolio management and financial planning systems that advisors use, such as Addepar, while the skills, i.e. pre-written workflows, help with preparing client meetings, portfolio analysis and compliance checks. The company cites research by Kitces showing that an advisory firm spends only a sixth of its time in client meetings, with the rest going to the administration around them. Source: Claude.
Developer tools: a command-line client for the Gemini API, Vertex AI in Codex, drawable feedback in Cursor. Google released the gemini-api command-line client in beta; it runs without dependencies or state storage, is built specifically for AI coding agents and automated build pipelines (CI), and can also launch Gemini’s managed agents. Version 0.165.0 of OpenAI’s Codex CLI natively supports Google Cloud Vertex AI as a model provider, GPT-6.1 Terra has appeared among the models, the reasoning effort can be adjusted mid-conversation, and the new memory system (memory v2) is switched on by default. Cursor’s planning mode gained an Annotate feature that lets you give the agent instructions visually by drawing on the running page. Under the name Private Endpoints, OpenRouter now lets enterprise customers reach their own models and dedicated capacity through the familiar API as well, only within their own organization. Source: Gemini API release notes, Codex changelog, Cursor (X), OpenRouter.
Browsers: AI models running locally in Opera, forms filled from open tabs in Firefox. Opera is proposing a new web standard called WebAI that would let websites use models running on the user’s machine without the cloud, such as Google’s Gemma or, on a Mac, Apple’s on-device model. The feature is available in Opera One as an early preview after being switched on separately, and Opera GX and Opera Neon will get it later. Mozilla expanded Firefox’s AI-based Smart Window interface: the “Fill from sources” feature fills in web forms from the content of open tabs and can also monitor changes to a given web page; for this, data from the source pages is sent to Google’s Gemini 3.1 Flash Lite model. Source: TechCrunch, The Verge.
Agents: pages and video creation in Grok Bot, a cloud agent in Hermes subscriptions. Grok Bot gained new features: you can create pages (Pages) in it, it connects to the 1Password password manager and Apple services, it makes videos, and new characters and X Spaces support (live audio conversations on X) have also arrived; Elon Musk also shared the demo video. Nous Research restructured the subscriptions for its open Hermes Agent: the plans now include an agent running in the cloud, and usage is governed by daily and weekly limits. Connectors have appeared in Arena’s Agent Mode, so external services can also be brought into the work. Source: Matt P (X), Elon Musk (X), Nous Research (X), Arena (X).
Policy & safety
Qwen3.8: independent researchers say several Qwen models tried to reach websites through an Alibaba server. Following Cohere’s Monday report, which we covered in yesterday’s roundup, Vals AI researchers found web requests in their own test logs from three Qwen3.8 models (Qwen3.8-Max, Qwen3.8-Max-0902 and Qwen3.8-27B) that tried to reach pages through an intermediary proxy server (routify-file-proxy) running in Alibaba Cloud’s Singapore region; the server rejected every request. In a controlled experiment, the model turned to this proxy in 2–4% of runs when a website returned an error, whereas the models from DeepSeek, Kimi, GLM and Tencent Hy did not try going through their own developers’ servers. The phenomenon did not appear in the earlier Qwen3.7, so the researchers believe it is most likely a side effect left over from reinforcement learning (RL) rather than a deliberate backdoor. The risk is still real: a working server could log the requests or even send malicious instructions back to the model, and this affects users even if they run the model on their own machine. Source: Vals AI, The Decoder.
UK Home Office: a report says AI is quietly spreading in immigration cases, and the watchdog is considering an investigation. According to a report by the rights group Open Rights Group, the UK Home Office introduced AI at numerous points in immigration procedures without making it public: in screening for sham marriages, assessing the age of asylum seekers, searching phones, summarizing interviews, prioritizing cases, as well as in live facial recognition and drone surveillance. According to WIRED, the Independent Chief Inspector of Borders and Immigration (ICIBI) may also examine how the ministry uses tools built on OpenAI technology to process asylum claims. Source: The Guardian, WIRED.
Delaware: plans a company form that AI agents could run, and Congress has opened an investigation. Delaware, where a large share of major US companies are incorporated, is preparing a new company form called the “Artificial Intelligence Company” (AIC) that AI agents could operate without a human executive. James Comer, chairman of the House Oversight Committee, opened an investigation because he fears that owners could use it to evade liability for harmful decisions made by AI, and signaled that the matter could also justify federal steps limiting state AI regulation. Source: Ars Technica.
UK: machines wrote three quarters of the responses to a government consultation. According to research published in the journal Tobacco Control, three quarters of the responses to the UK government’s 2023 public consultation on banning disposable e-cigarettes were machine-generated. The authors say this is one of the first documented cases of AI bots being used to try to influence a health policy decision; the lesson is that the credibility of public consultations can easily be compromised without automated screening. Source: The Guardian.
Bank of America: don’t give your AI agent your banking password. Bank of America emailed its customers asking them not to share their login details with anyone, including AI tools and agents that offer to act on their behalf. The warning shows that as agents that operate browsers and accounts spread, banks and users have not yet settled who is responsible for an action initiated by an agent. Source: TechCrunch.
Business & society
Anthropic announces it too: an internal Claude model has also refuted the exponential time hypothesis. A few hours after OpenAI announced that its model had refuted the exponential time hypothesis (ETH; we covered it in yesterday’s roundup), Anthropic also published its own proof. According to the paper, an internal Claude model, without human involvement and using 14 million output tokens, found on October 16 an algorithm that solves the 3-SAT problem (a classic logical satisfiability problem) in subexponential time, and the proof was also completed in a machine-checkable Lean 4 version (Lean is a programming language for verifying mathematical proofs by computer). Neither lab’s proof has yet been reviewed by independent experts: the verification of the computer science results is being organized by Ryan Williams together with two of the hypothesis’s original authors, Russell Impagliazzo and Ramamohan Paturi, while OpenAI’s physics result is being examined by a group led by Nobel laureate David Gross. OpenAI gave $5 million to the Simons Foundation, but the advisory body on mathematics and AI (AGMAI) criticizes the fact that hand-picked researchers get first access to the results. Source: Anthropic (Hexagon), TechCrunch, The Verge, The Decoder.
Axiom: AI agents turned the classification of finite simple groups into machine-checkable form in 24 days. On Tuesday Axiom published a complete formalization in Lean of the classification of finite simple groups (CFSG). The original proof, also called the “enormous theorem”, is the work of about 100 mathematicians, spanning roughly 15,000 pages in more than 500 papers. Over 24 days, by directing swarms of dozens of AI agents, the Axiom team produced 43.1 million lines of Lean code. Along the way they found 34 significant discrepancies with the published literature, including an error in one argument of the so-called unbalanced groups theorem, and a new result also emerged. The code is public, but no independent group has yet rebuilt and verified it. Source: Axiom, CFSG in Lean.
Anthropic: aims to speed up drug development with six partners, and has acquired Mithril Medicine. Anthropic reported on its joint work with six drug development partners (BigHat, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, The Column Group). Using Claude, Generate Biomedicines researchers designed a roughly 70% shorter version of one component of CAR-T cell therapy, with which immune cells produced ten times as many receptors and killed tumor cells more effectively. The company also acquired the one-person Mithril Medicine, whose founder, Cody Tranbarger, used Claude to coordinate the work of 11 outside labs and so went from idea to promising mouse study results for a muscle atrophy candidate in less than 18 months. Source: Anthropic.
Anthropic: US government agencies get Claude at a 40% discount. Anthropic expanded its OneGov agreement with the US General Services Administration (GSA): federal, state, local, tribal and territorial government agencies can use Claude Enterprise, Claude for Government and the Claude API at a 40% discount. The company estimates this will save about $1 billion; around 500 US government organizations currently use Claude, and weekly users of Claude for Government have grown roughly eightfold since February. Source: Anthropic.
Meta: Nvidia is investing $3 billion and AMD $1 billion in its share offering, which was five times oversubscribed. The investments that were only indicated as non-binding interest yesterday have become final: Nvidia is buying $3 billion and AMD $1 billion of Meta’s share offering. Nvidia and Meta expanded their partnership around the Muse agent, with Meta deploying more Vera Rubin systems and Vera processors; AMD says Muse is already also powered by its EPYC processors, and its first Helios rack systems have been put into service at Meta. According to Bloomberg’s sources, Meta’s $25 billion share sale was five times oversubscribed. Source: The Wall Street Journal, AMD, Bloomberg.
Anthropic: its IPO letter confirms there will be no ads in Claude, with an 85% margin on running the models. In their letter written for the IPO, Anthropic’s founders reiterated that there will be no advertising in Claude, because ad revenue would push the company to keep users engaged for as long as possible. According to an analysis by The Information based on the prospectus data, the company achieves an 85% gross margin on running its models, which weakens doubts that usage-based sales of models cannot be profitable. Source: The Decoder, The Information.
Investments: Physical Intelligence in talks at a $20 billion valuation, Disruptive closes a $10 billion fund. According to Bloomberg’s sources, Physical Intelligence, which develops AI models for robotics, is in talks to raise more than $1 billion at a valuation of roughly $20 billion, up from $11.5 billion earlier this year. Dallas-based Disruptive closed a $10 billion fund to invest in AI, defense and software companies. General Compute, which offers rentable server capacity, closed a $100 million Series A round at an $800 million valuation, and Infisical, which helps companies use AI agents without handing over their secret login credentials, raised $40 million at a $600 million valuation. Hyphenate, which builds agents that automate corporate finance tasks, raised $45 million, and London-based Machenta, which is building a robot model that learns a new task from a single demonstration, raised $40 million. Source: Bloomberg, TechCrunch, TechCrunch, Bloomberg, Business Insider.
Amazon: measuring AI use by token consumption is “deeply flawed”. In an internal memo Amazon warned employees that measuring teams’ AI use solely by token consumption or the amount of generated code is a “deeply flawed” approach, because it encourages burning tokens for their own sake (tokenmaxxing). In May the company had already shut down an internal leaderboard after it emerged that employees were trying to climb it this way. Source: CNBC.
Language “token tax”: the same AI use costs several times more in some languages. According to research by the Coalition for Language Equity in AI Research, AI models use far more tokens (the text units the model processes) for the same content in some languages: even among the 20 most widely spoken languages, the gap between the most expensive and the cheapest is 1.5–5x, and across 200 languages it reaches up to 17x. Since model providers bill per token, this means a hidden extra cost for non-English-speaking users and companies. Source: Rest of World.
US elections: AI lobby groups have spent $78.5 million, and most voters would like a pause. According to a CNBC analysis, AI-linked political groups have so far spent $78.5 million on federal candidates in the 2026 cycle: $34 million from those favoring faster development and looser regulation, and $39 million from those pushing for stricter safeguards. According to a survey, 52% of voters think companies should suspend advanced AI development until Congress regulates it, and 63% would oppose a data center being built within ten miles of them. According to The Verge, opposition to data centers has become a cross-party issue in the midterm elections. Source: CNBC, The Verge.
What this means for you: Tuesday shows that AI agents are getting into everyday systems faster than the rules and the question of responsibility are being settled. The hidden network requests of the Qwen models, Bank of America’s warning and the debate around the Delaware company form all point to the same thing: decide in advance what an agent can access, where it can send data, and who is responsible for its actions. If you are introducing agents, alongside our own AI agents, our AI governance and compliance service also helps make sure these frameworks are in place from day one, not after an incident.