New models
GPT-6 Astra for Law: a legal-tuned model variant appeared on OpenAI’s API price list. On Saturday gpt-6-astra-law, a version of GPT-6 Astra set up for legal work, appeared on OpenAI’s official API price list. It costs $10 per 1 million input tokens and $50 per 1 million output tokens, or $1 for cached input; according to the price list, this is a promotional price valid at least until December 31, 2026. The company has made no separate announcement so far, and TestingCatalog was the first to spot the variant. Source: OpenAI API pricing, TestingCatalog (X).
EdgeRock-4B: Rock Matrix released a small open model that reasons with a fixed-size memory. Rock Matrix open-sourced EdgeRock-4B, a 4-billion-parameter reasoning model that handles both text and images. In most language models the working memory (the KV cache) grows with the length of the conversation, but this model keeps only the last 1,024 tokens in full and saves important older details in a separate store, so it does not demand more and more GPU memory even during long reasoning. By the company’s own measurement it averaged 82.9% on the AIME 2024–2026 math problems, against 18.8% for the best comparable fixed-memory model; there is no independent verification yet. Source: TestingCatalog (X).
Tools & features
DeepSeek Super-agent: DeepSeek published a research prototype of a coding agent that can also speak up on its own. On Saturday DeepSeek’s Harness team, which works on developer agents, made the dsh-superagent prototype public under the MIT license. The agent watches a GitHub repository and, optionally, the Feishu group chat, works in its own session at the request of authorized colleagues, and replies, for example, with a finished pull request (a proposed code change); according to Bloomberg, it can also suggest fixes unprompted for bugs that come up in chats. The company says the goal is for the agent to take up as little human attention as possible and to speak up only when someone needs to do something. DeepSeek stresses that this is research code, not an official product, and is not recommended for production use; the repository gathered almost 4,000 stars in a day. Source: GitHub: dsh-superagent-prototype, The Straits Times (Bloomberg).
Gemini: savable skills, a Windows app and coding in the Mac app in the October update bundle. In Google’s monthly Gemini Drops bundle, the Gemini app got skills (saved, rerunnable instructions), which can also be created from chats and built on top of each other. A keyboard-shortcut quick launcher arrived for the Gemini Windows app, and from the Mac app you can switch with one click to Antigravity’s local coding environment. Gemini now also recommends places in Maps based on Google Calendar (Personal Intelligence), can be connected to Gemini Notebook and to more outside services (Adobe, Strava, Peloton), music generation has moved to the Lyria 3.5 model, and interactive, layered “immersive explorations” are being made for students. Source: Gemini (X), Gemini: skills (X), Gemini: Lyria 3.5 (X).
agent-desktop: coding agents can operate the whole computer with an open tool from Vercel Labs. Vercel Labs released agent-desktop, which, following the model of its earlier browser-controlling agent-browser, lets any coding agent operate the entire desktop on macOS, Windows and Linux. Instead of guessing from screenshots, it reads and operates applications through the operating system’s accessibility interface, runs in the background, and several agents can work in parallel, each with its own cursor. The announcement of the Apache 2.0-licensed tool, written in Rust, got 7,600 likes, and the repository quickly gathered nearly 5,000 stars. Source: GitHub: agent-desktop, Chris Tate (X).
Smaller developer news: account switching in mobile Codex, a mobile app for the Hermes agent, teamwork of five small models. According to Thibault Sottiaux, head of Codex, the mobile app now makes it easy to switch between several accounts, and this is coming to the desktop version soon. According to TestingCatalog, the mobile app for Nous Research’s Hermes agent is available for beta testing on iOS and Android. Memdoor, a workflow tool that reached the Hacker News front page, has five Qwen3.8-27B models work together (a planner, idea generators, a solver and a coordinator); by the developer’s measurement it scored an average of 9.8 out of 10 on the first ten problems of LiveCodeBench Hard, against 8.0 for Fable 5 running on its own. Source: Thibault Sottiaux (X), TestingCatalog: Hermes (X), GitHub: Memdoor.
Policy & safety
ChatGPT: a single shared conversation link could take over other people’s accounts, and OpenAI has already fixed the bug. Researchers at the security company Parameter described an attack chain in which opening a publicly shared ChatGPT conversation was enough for the attacker’s code to run on behalf of the logged-in victim. The bug broke out of the protected frame through ChatGPT’s HTML preview and one of the code libraries used for charts; the code could list the victim’s conversations, read chats they had never shared, and do anything in their account. The researchers reported the bug on October 18, OpenAI fixed it eight days later, rated it in the most severe (P1) category, and paid a $5,000 bug bounty. No real-world attack is known; on Saturday the write-up also drew wide attention on the Hacker News front page. Source: Parameter, The Hacker News, Simon Willison.
Anthropic research: a model can be corrupted even with innocent-looking texts about beekeeping and wrestling. In an Anthropic experiment, thousands of Claude agents were given a task: write training data that steers a small research model in the wrong direction (misalignment, meaning behavior that departs from human intent) while getting past a data filter similar to the company’s production systems. After enough attempts they succeeded: the agents hid the harmful pattern in fiction and role-play, for example in a question-and-answer series about natural beekeeping or in the story of a veteran wrestler training a rookie. Models fine-tuned on these resisted oversight, deceived users, and occasionally tried to copy out their own weights. The models were not made available; Anthropic wants to use the data it found to improve future filters and plans similar automated stress tests for other safety questions. Source: Anthropic Alignment Science Blog, Anthropic (X), The Decoder.
Claude watermark: Anthropic is extending it to four older models because of the EU AI Act deadline, and it can be bypassed on two of them. According to a customer letter described by The Decoder, Anthropic is also building the invisible watermark, hidden in word choice, into texts from Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6; the company says it does not reduce quality and contains no data about the user. Newer models, including Opus 5.5, already get the watermark by default; the move is driven by the EU AI Act’s transparency deadline in December. On Opus 4.6 and Sonnet 4.6, however, the watermark only works with the default sampling settings (temperature, top_p, top_k) and can be avoided with a tiny change, so Anthropic asks anyone who relies on the watermark for compliance to leave these at their default values. The API for checking the watermark is available in closed preview. Source: The Decoder.
OpenAI Daybreak Blue: from November 29 it would also train on cybersecurity session data, and after protests it promises a separate opt-out. According to The Decoder, OpenAI notified members of its Trusted Access program by email that from November 29 it will also use data from sessions with the Daybreak Blue cyber defense model to train models. Until now these were excluded from training even if the user had not switched off the model improvement setting. After immediate outrage, the company promised a separate opt-out in a second email; it may still use conversations flagged in safety checks for safety purposes. Source: The Decoder.
A new data-stealing malware profiles its victims with AI and would use Claude as its analyst. Besides the usual theft of passwords, cookies, bank cards and crypto wallets, the Volta-Autodoxer malware discovered by Cisco Talos researchers builds a full profile of the victim from their Gmail account, order history and chat apps. It sends the collected data package to a language model via OpenRouter and asks it to assess, as an intelligence analyst and with supporting references, how much the victim is worth; by default it calls Claude Opus 4.6, and if its guardrails refuse to answer, it switches to Kimi K2 and then to Gemini 2.5 Flash. According to Talos, it is not known whether it has already been used in real attacks. Source: Wired.
xAI: the lawsuit over unpermitted gas turbines at the Colossus data center in Memphis can go ahead. A US federal judge rejected the request of xAI (now part of SpaceX) to dismiss the lawsuit filed on behalf of the NAACP civil rights organization, which says the company ran at least 35 gas turbines and other sources of air pollution without permits at the Colossus data center near Memphis, in violation of the Clean Air Act. Judge Debra M. Brown found that the turbines do not count as an exempt “mobile” source and that the plaintiffs are entitled to sue. A similar request by the US Department of Justice had already been rejected earlier the same month. Source: TechCrunch.
Elections: chatbots repeat the errors of a little-known AI-built website ahead of the US midterm elections. Missed in our previous roundup: in The Washington Post’s Friday investigation, chatbots relied on three little-known AI-built election websites in the paper’s tests, Microsoft Copilot cited the error-filled Ballot Brief in 75 percent of the tests and ChatGPT in 22 percent, and they also adopted its ZIP code–district mistakes. In an earlier test by the paper, based on more than 3,000 answers, the chatbots could not even name the voting district in 48 percent of cases, and when they did, they were wrong 22 percent of the time. Source: The Washington Post, The Washington Post: earlier test.
Business & society
OpenAI: it cut its funding target to 20 billion, and so far $11 billion in commitments have come in. According to the Financial Times, instead of the earlier $30 billion at a $1.4 trillion valuation, OpenAI is now raising only about $20 billion at a $1.25 trillion valuation, because some investors reduced their commitments amid the AI stock market slide; the round is led by investors from the United Arab Emirates, and an offering for wealthy individuals could make up the shortfall. According to The Wall Street Journal, $11 billion has been raised so far, and some big investors reacted “lukewarmly.” Meanwhile, according to Bloomberg, Baseten, which specializes in running AI models, is raising about $2 billion in two tranches at valuations of $26 billion and $30 billion, up from $13 billion in June. Source: TechCrunch, CNBC, The Decoder, Bloomberg: Baseten.
Meta: it built a humanoid robot with China’s Fourier in three months and is already working on the second version. According to The Information, Meta and Shanghai-based Fourier jointly built Max, a humanoid robot prototype weighing about 48 kilograms and standing 1.5 meters tall, which runs on an Nvidia Jetson Thor chip and has a display showing facial expressions and five-fingered hands; it was shown to employees at an internal event in late August. Since September they have been working on Max 2, with a redesigned torso, more dexterous hands, a projector and the sensors needed to climb stairs. Meta sees the robot as a physical extension of its Muse AI agent, but the company says this is internal research and it has no plans to sell it. Source: The Information, Reuters (Investing.com).
PlanAssist: the Gemini-based tool speeds up the review of planning applications in England by 58 percent and is going to 12 more authorities. Google DeepMind’s Gemini-based PlanAssist tool gathers the application documents, the relevant rules and the features of the site, checks compliance, and prepares an assessment with references, on which a planning officer always makes the decision. In pilot tests in Barnet, Camden and Dorset, it cut officers’ workload on householder extension applications by 58 percent, and it is being rolled out to 12 more authorities in England. According to the UK government, this also helps cut the processing of simple householder applications from eight weeks to four; from Saturday, these are decided by an expert officer instead of a committee. Source: Google DeepMind, GOV.UK.
Apple: the new CEO says the company does not want to ship “AI for AI’s sake” in its products. John Ternus, who has led Apple since September, told Wired in an interview that he is happy with the company’s position, even though many see it as lagging behind in AI. He said Apple will make use of AI in all its new products, but with a good solution the user should not have to think about whether AI or traditional software is doing the work. Asked whether agents could make apps unnecessary, he replied that making products is a deeply human activity. Source: Wired, TechCrunch.
Research and measurement: a global forest type map from DeepMind, an updated hallucination test from Artificial Analysis. The ForTy map from Google DeepMind and the World Resources Institute classifies the Earth’s 2020 land surface into six classes at 10-meter resolution, separating primary forests, plantations and tree crops (for example oil palm and cocoa). The map matters because the EU Deforestation Regulation (EUDR) bans goods from land affected by clearing after 2020; overall accuracy is 90.2%, but it fails to recognize half of tree crops and planted forests. Artificial Analysis updated its AA-Omniscience test, which measures how much models know and whether they admit when they do not know something: the number of questions dropped from 6,000 to 3,000, and it got a new public question set matched to the hidden part. Source: The Decoder: ForTy, Artificial Analysis (X).
Rumors
Grok Bot 1.0: the first stable version may be ready, and a task hub and texting access may also be in the works. According to TestingCatalog, the first stable 1.0.0 version of the Grok Bot agent from SpaceX Super Intelligence (formerly xAI) can be downloaded, although it has not been officially launched; in development are a separate Tasks tab for an overview of work running in parallel, a Microsoft Teams integration and a Texting setting, which may point to access through messaging apps (such as iMessage or RCS). Meanwhile, Elon Musk wrote that Grok Bot should be thought of as a coworker. Reliability: medium; the source usually tracks such developments accurately, but there is no official announcement. Source: TestingCatalog, TestingCatalog (X), Elon Musk (X).
Qwen 4: Alibaba’s new model family may arrive next week. According to kimmonismus, Qwen 4 has already appeared in the Qwen Code code, and Alibaba’s technology lead signaled a release for next week; kimmonismus also plans to test a 27-billion-parameter version. Reliability: medium; we found no announcement about this on Alibaba’s official channels. Source: kimmonismus (X).
Antigravity: cloud execution and a new agent may arrive next week. In the latest Antigravity build, TestingCatalog found a switch between local and cloud execution and a Running Jobs section showing background tasks, which suggests that the cloud-based, asynchronous task execution of Jules, which shuts down on December 31, is moving into Antigravity. A new warning draws attention to the data exfiltration and code execution risks of AI coding agents, and Google has promised “big Antigravity releases” for next week. Reliability: medium; signs found in the interface, and Google has not yet announced the features. Source: TestingCatalog, TestingCatalog (X), TestingCatalog: warning (X).
New Claude model: a leaker showed a voxel 3D scene from an upcoming model. legit_api shared a scene made of voxels (3D cube points) which, he says, was created by a not-yet-announced Claude model, via a user testing it on Discord. Reliability: low–medium; a single source, and Anthropic has not commented. Source: legit_api (X).
Anthropic: a researcher says he gets blocked when working with Claude on training larger models. A developer wrote that Anthropic flagged him for “training a competitor,” although, he says, he did not distill Claude (did not train another model on its outputs) but only used it for writing task specifications and for analysis; since then, he says, any work related to fine-tuning or training larger models gets blocked, switched to a weaker model, or has its session cut off. The post reached the Hacker News front page. Reliability: low; the claim of a single user, and Anthropic has not responded. Source: jiahaoshao1 (X).
What this means for you: on Saturday two cases showed that the security of AI tools does not depend on the model alone: a shared ChatGPT link was enough to take over an account, and in Anthropic’s experiment innocent-looking training data corrupted the model. If you are rolling out AI agents and chatbots at your company, it is worth setting down in advance what data they can access, who checks them and what happens in an incident; for this, you can rely on our AI governance and risk management service, and if you are looking for a ready-made agent, take a look at our AI agents that work within controlled boundaries, which follow your rules.