New models
GPT-6 Rosalind: OpenAI releases a GPT-6-based life sciences model for approved research groups. On Friday OpenAI released the GPT-6 Rosalind Research model, which brings GPT-6’s reasoning abilities to multi-step life sciences work (biology and drug research), with safety limits that make biological misuse harder. Only research organizations approved in OpenAI’s trusted access program can use it, in the Work and Codex interfaces and, with separate approval, in the API as gpt-6-rosalind; it is not available in the regular chat tab. In the API, 1 million input tokens cost $15 and 1 million output tokens $75, and in enterprise workspaces it starts with a two-week admin preview that is switched off by default. The lower Discovery access tier stays on the earlier Rosalind 5.5 model. Source: OpenAI API changelog, OpenAI: model release notes, OpenAI: GPT-Rosalind.
Claude Opus 4.8: fast mode is being retired and will remain only in Opus 5.5 after November 30. On Friday Anthropic deprecated the fast mode of Claude Opus 4.8 (the setting for quicker responses) and will shut it down for good on November 30: after that, requests sent to the claude-opus-4-8 model with fast mode will return an error and will not automatically fall back to normal speed. Anyone who wants to keep the speed-up has to move to Opus 5.5; Claude Code also warns about the change. Source: Claude Platform release notes, Claude Code changelog.
Small open models: SenseTime’s 9-billion-parameter model runs layers twice, and Nvidia releases a noise-cleaning speech model. China’s SenseTime released the SenseNova-Loop-9B-Preview model under the Apache 2.0 license; it runs 24 of its 32 layers twice in a row with the same weights, so it does more computation while staying small. It accepts text and images, works with long documents and multi-step agentic tasks, and according to the company’s own measurements it beats Alibaba’s three times larger Qwen3.5-27B model on several office agent tests. Nvidia’s NemotronLabs RE-USE, an open model with just 9.6 million parameters, cleans speech recordings of background noise, echo and other distortions in several languages. Source: Hugging Face: SenseNova-Loop-9B-Preview, Hugging Face: RE-USE, NVIDIA AI (X).
Specialized models from startups: a model that learned to teach, a cheap financial research agent and a video model inspired by living neurons. humans& released the &Tutor (Ampersand Tutor) model, which was taught to teach by training it against an in-house model that simulates a student; the company says it achieved the best result on the TutorMoments test, built from real tutoring situations, and anyone can try it through the API. Samaya AI’s financial research model Sterling (30 billion parameters, 3 billion of them active, built on Nvidia’s Nemotron 3.5 Lightning model) scored 58.9% on the company’s own FrontierFinance test versus 55.9% for Claude Fable 5.1, at $0.61 per question instead of $6.89. The Biological Computing Co. says its image-to-video model mischief 1 is five times faster than its base model because it uses efficiency techniques borrowed from studying living neurons grown in the lab; it is available on OpenRouter and fal for half a cent per second. Source: humans&, Samaya AI, The Biological Computing Co. (PR Newswire), OpenRouter (X).
Tools & features
Claude: the $200 plan is renamed Ultra and loses its five-hour limit. Anthropic renamed its Max plans: the $100-a-month Max 5x is now simply Max, and the $200-a-month Max 20x is now Ultra; prices do not change and subscribers do not need to do anything. On Friday the five-hour usage window was removed for verified Ultra accounts, starting with most existing subscribers; for fraud prevention, newer accounts get a temporary limit until they are verified, and receive an email when it is lifted. Weekly limits stay the same. Source: Claude (X), Boris Cherny (X).
OpenAI Private Intelligence: OpenAI bundles its business privacy tools into one package, with encrypted processing as a new element. OpenAI unified its business data protection commitments under the name Private Intelligence: company data (both inputs and outputs) stays owned by the customer, the admin decides what agents can access and when approval is needed, and Zero Data Retention can be requested for eligible API workflows. The new element is Private Inference, launching in preview, which uses confidential computing to protect requests and responses even while they are being processed. OpenAI did not publish prices or the list of eligible models; access can be requested through sales. Source: OpenAI, OpenAI (X).
Gemini API: computer use is now generally available. Google released the Gemini API Computer Use tool for production use (GA) in the Interactions and GenerateContent APIs: based on screenshots, the model returns clicks and keystrokes for browser, mobile and desktop agents, while the actions themselves are carried out by the developer’s own system. The tool has no separate fee; it is billed at the selected model’s usual token price. Source: Gemini API release notes, Gemini API: Computer Use.
Google: the Jules coding agent is shutting down, and its capabilities move into Antigravity. Missed in our previous roundup: on Thursday evening Google announced that it is retiring Jules, the coding agent that works autonomously in the background, and will build what it learned into the Antigravity agentic development environment, so that remote background task execution and agent-centered development are available in one place. On Friday Antigravity also became available in Android Studio, where it can use Android development’s own tools (emulator, build system, logs). Source: Jules (X), Antigravity (X), Android Developers Blog.
Apple Intelligence: coming to China in December, with Alibaba’s Qwen model. Apple announced in its Chinese newsroom that the first Apple Intelligence features will launch in China with iOS 27.3: notification summaries, live translation, writing tools for rewriting and proofreading, visual intelligence and removing objects from photos. The beta arrives on November 4, the final release in December. The new AI-powered Siri is not coming, but some features will offer Alibaba’s Qwen chatbot, similar to how ChatGPT was built in elsewhere. Source: The Verge.
Kimi 3D Make: Moonshot’s assistant builds 3D models with Blender and CAD programs. The Kimi assistant from China’s Moonshot AI received a new feature called 3D Make: by connecting to Blender and CAD design programs, it creates 3D models, assembles scenes and develops games. It is available on kimi.com and in Kimi Work. Source: Kimi (X).
Creative tools: Runway’s agent opens files and runs code, live web search in Lovable apps, voice settings in Grok Bot. Runway Agent no longer just generates: it opens files, runs code, and from a single reference creates a brand launch kit, product photos, ad variations and motion graphics. Apps built with Lovable can use real-time, source-backed web information through a new Perplexity connector. According to TestingCatalog, voice and language settings are now available in Grok Bot, so individual bots can speak with different voices and in different languages. Source: Runway (X), Lovable (X), TestingCatalog (X).
Developer tools: writing outside the working directory can be blocked in Claude Code, many tool calls in one request in Factory Droid 1.0, connectors in Arena’s agent mode. In Friday’s Claude Code 2.1.314 release, a setting can prevent the agent from writing files outside the working directories, and in VS Code and the Code tab of Claude Desktop, idle conversations longer than 200,000 tokens are automatically summarized. Missed in our previous roundup was Factory’s Thursday release of Droid 1.0: the coding agent can write a program in a single model request that runs many tool calls, and the company says it scores better on the Terminal-Bench 4 test and is up to 24% cheaper than solutions compared using the same models. In Arena’s agent mode, Vercel, Linear, GitHub and other tools can be connected in a single task, and on GitHub the most new stars went to claude-mem, which gives memory across sessions, and OpenMausBot, an open-source Grok Bot alternative that gives bots their own virtual machine. Source: Claude Code changelog, Factory, Arena (X), GitHub: claude-mem, GitHub: OpenMausBot.
Policy & safety
A malicious Claude Code extension landed on Cursor’s marketplace and may have affected up to 5,000 developers. On Friday Cursor said that on Thursday, for nearly 11 hours (between 10:27 and 21:35 Madrid time), counterfeit versions 2.1.312 and 2.1.313 of the Claude Code VS Code extension could be downloaded from its extension marketplace for Windows and macOS. The malware sent out machine information and the names of environment variables, installed the attacker’s SSH key and tried to set up a persistent, outward-facing SSH tunnel; according to Cursor, this failed on most machines because it lacked administrator rights, and the attacker’s server is no longer running. Cursor took the marketplace offline, audited and rebuilt it, and asks everyone to update to version 2.1.314. Anthropic says its own systems were not breached and the extension distributed through official channels is not affected. For those affected, removing the extension is not enough: the installed key and the startup scripts must be deleted separately, and credentials must be changed. Source: Cursor, Cursor (X), Felix Rieseberg, Anthropic (X), Ars Technica, Microsoft Security.
A shared signal in robots.txt: OpenAI, Google, Microsoft and Anthropic will all respect a site’s wish not to be used for AI training. The four companies announced that they will adopt a shared signal (policy token) proposed at the IETF, the body that develops internet standards, with which websites can tell every participating AI company at once in their robots.txt file that they do not want their content used for model training; until now this had to be set up separately for each company. The proposal was made by Anthropic’s standards specialist Kevin Kelley. Microsoft offers separate opt-outs for training and for models citing the site as a source (grounding), and sites remain in Bing results regardless. Meta said it supports the initiative but did not say whether it will actually comply. Source: OpenAI, The Verge.
ChatGPT: fewer weapons details for teens, it can end conversations fixated on violence, and a safety report every six months. OpenAI tightened violence prevention in ChatGPT: teenagers get less practical information about weapons and security systems even when the question seems harmless, and all users get fewer details about past mass attacks. ChatGPT can end conversations fixated on violence and, in repeated cases, suspend the account; the company is using more accurate age estimation, and it will also notify parents if their child’s account is suspended for this reason. From early 2027, OpenAI will publish a community safety report every six months. Source: OpenAI, Reuters (CNA).
UK competition authority: an investigation into Meta’s terms, under which it would also use company websites for AI. The UK competition authority (CMA) opened a case to establish whether Meta is abusing its dominant position to impose unfair terms on publishers and advertisers. Under the disputed new commercial terms, Meta could crawl and use the content of company websites for its own products, including AI, without a meaningful choice or way to opt out. Meta has paused the rollout pending industry consultation; the CMA is gathering evidence until January 2027 and has not yet found any infringement. Source: CMA (GOV.UK), Reuters.
Copyright: Wikipedia’s operator says training can be fair use, but reproducing articles word for word is not. On Friday the Wikimedia Foundation filed a brief (amicus brief) in the lawsuits of The New York Times and other publishers against OpenAI and Microsoft, without siding with either party: in its view, courts should protect training large language models on copyrighted works, but not chatbots reproducing those works word for word. An overly black-and-white ruling, it argues, would either take useful AI tools away from Wikipedians or endanger the open web. On the same day OpenAI quoted briefs from creators, researchers, accessibility organizations and lawyers arguing in favor of fair use. Source: The Verge, OpenAI.
CrowdStrike: Chinese state hackers used an AI tool built on a DeepSeek model to attack more than 200 organizations in 11 days. Missed in our previous roundup: according to an analysis CrowdStrike published late on Thursday, a group with Chinese state ties used an AI-powered attack tool called ARTEX, including DeepSeek and another language model, to target more than 200 government and business organizations worldwide in 11 days, breaching at least nine. According to NBC News, the case shows that with AI, attackers can work at a scale that used to require many more people. Source: NBC News.
Pope Leo XIV: AI must be disarmed; machines are tools, not persons. In a speech on Friday, Pope Leo XIV called on UN member states to take concrete steps toward banning lethal autonomous weapons at their meeting in Geneva next month. He said decisions affecting human life, freedom and the use of force must always remain under real, accountable human oversight, and warned that it would be dangerous to grant rights or legal personhood to AI systems. Source: The Verge.
Business & society
AI spending and the stock market: tech stocks lost nearly $1 trillion in a week, Intel and Reddit fell by record amounts. The stock market slide driven by doubts over the return on AI investments continued on Friday: according to the Guardian, Nasdaq tech stocks lost nearly $1 trillion in value in a week and about $2 trillion in a month, falling more than 3% on Friday. Intel’s share price plunged 29% after a weak forecast (an Israeli factory shutdown, a weak PC market), its worst day since the 1970s, cutting the value of the US government’s stake by about $12 billion; Reddit fell 29%, and its CEO said Google’s AI Mode “doesn’t drive traffic for anybody”. According to Bloomberg, Alphabet, Meta, Microsoft, Amazon and Oracle together have more than $3 trillion in AI infrastructure commitments (Alphabet more than $1 trillion of it, Meta about $800 billion), and S&P cut Meta’s outlook to negative because of heavy capital spending and negative free cash flow. Source: The Guardian, CNBC: Intel, CNBC: Reddit, The Decoder, Bloomberg, CNBC: Meta.
Anthropic: may receive a “junk” rating from credit agencies after its IPO. According to the Financial Times, Blackstone, the lead investor in a $63 billion structure financing Anthropic’s chips, told investors that after going public the AI lab could receive a credit rating in the upper part of the band below investment grade, the so-called “junk” range. This could make financing Anthropic’s huge compute expansion harder and more expensive. Source: TechCrunch.
Cohere: investing more than 4 billion Canadian dollars in Canada by 2031, with 1,500 new jobs. A day after its record funding round, Canada’s Cohere announced that its Canadian program, launched in 2024, will involve more than 4 billion Canadian dollars of research and development and AI infrastructure investment and more than 1,500 new jobs by 2031; at least 1,000 of them will be research, engineering and product development positions. The company is also working with the University of Toronto and the University of Waterloo. Source: Cohere, Cohere (X).
Project Suncatcher: an AI model answered for the first time from Google’s orbiting, TPU-equipped satellite. Google announced that the prototype satellite of Project Suncatcher, which prepares the way for data centers in space, ran a request for the first time: the Gemma 4 model running on the onboard TPU wrote a haiku, which came back to Earth during a pass over Svalbard. The satellite has four TPUs, but only one can run at a time; the next step is studying thermal load and the effects of radiation, and in 2027 two more satellites will launch with more advanced chips and optical links between satellites. Source: Google, TechCrunch, Sundar Pichai (X).
Defense AI and investments: Salesforce buys Rune for $375 million, Google in talks about Golden Dome. According to Resilience Media’s sources, Salesforce is buying Rune Technologies, which uses AI to organize front-line supply chains, for $375 million in cash and stock, to strengthen Missionforce, its year-old business unit for military customers. According to Bloomberg, Google is in early talks with the Pentagon about helping secure the data flowing into the command software of the Golden Dome missile defense program. Powerlattice, which develops chips that reduce the power needs of AI processors, raised $86 million. Source: Resilience Media, Bloomberg, GeekWire.
AI agents in commerce: only a few hundred orders at DoorDash, invisible ads for machines. DoorDash CEO Tony Xu said on the quarterly earnings call that only “several hundred” orders in total have so far come through outside AI agents such as Meta Muse, while the company delivers millions every day. Meanwhile, according to The Information, at least 11 publishers, including Time magazine, have started showing ads that only AI agents read and human visitors cannot see; Perplexity already blocks these from influencing its tools. Source: TechCrunch, The Information.
Sam Altman: it is bad for memory to sit in home computers, and OpenAI’s math results will keep coming. On the Cheeky Pint podcast, the OpenAI chief said that given the tight DRAM supply it is “pretty bad” for memory to sit in users’ machines working only 5% of the time, so instead of running models locally he thinks it makes more sense to use data centers as fully as possible. He also said OpenAI has spent hundreds of millions of dollars on math research; he understands mathematicians flooded with a series of AI results, but they will not stop, and he hinted that more discoveries may be in the pipeline. Source: The Verge: memory, The Verge: math.
Research: a Claude agent chose a gene therapy on its own and carried out brain surgery on a mouse. According to Gabriel Ocana-Santero, head of the startup Exin Therapeutics and a doctoral student at the University of Oxford, an AI agent built on Claude Sonnet 5 chose the treatment from 19 possible gene therapies, as well as the brain region and the injection coordinates, in a mouse model of epilepsy, and then carried out the procedure by controlling a motorized surgical device. Every irreversible step was approved by a licensed veterinarian, who also held an emergency stop; during surgery the agent detected a calibration error and refused to go beyond its safety limits. Meanwhile, according to Epoch AI data, 17% of biology preprints (papers shared before peer review) already report using language models. Source: The Decoder, Epoch AI (X).
Rumors
GPT-6.2 “Bel”: leakers say a new top OpenAI model is approaching. According to community leakers, OpenAI’s next big model could be GPT-6.2, code-named “Bel” (Betelgeuse), built on a new pretraining run; kimmonismus also passed on the report but added that he cannot confirm it. Reliability: low; claims from anonymous accounts, and OpenAI has not commented. Source: kimmonismus (X), taker_of_whizz (X).
Google Omni Flash 1.2: a new video model with 4K output may have surfaced. According to chetaslua, Google’s Omni Flash 1.2 video model is already available on a limited basis, with 1080p and 4K resolution, frame interpolation, reference images and video extension; he also shared a sample video. Reliability: low–medium; Google has announced nothing. Source: chetaslua (X).
Antigravity: a background tasks tab and scheduled runs may arrive after Jules. According to TestingCatalog, a “Running” tab for tasks running in the background is being prepared in Antigravity’s settings, and after Jules shuts down, advanced scheduled runs may also move here; traces of an agent called “Chief of Staff” were spotted earlier as well. Reliability: medium; based on signs found in the interface, not confirmed by Google. Source: TestingCatalog (X), TestingCatalog: Jules (X).
What this means for you: on Friday a routine update showed how vulnerable AI tools running on developers’ machines are: a counterfeit Claude Code extension could install itself automatically on thousands of machines for hours. It is worth setting down where agents and extensions can come from on company machines, with what permissions, and who checks them; for this, you can rely on our AI governance and risk management service. Meanwhile models and plans keep changing, with Opus 4.8’s fast mode, for example, disappearing in a month, so if you are planning AI costs, take a look at our transparent pricing plans, ideally before tying yourself to a single provider.