New models
MiniMax M3.1 Flash: the anonymously tested "Space Bunny" model launches as a cheap, fast agentic model. China's MiniMax released M3.1 Flash on Monday and revealed that it was the model tested anonymously under the name "Space Bunny", which processed more than 200 trillion tokens in the meantime. According to the company it is roughly three times faster and 70% cheaper than its predecessor M3, priced at $0.15 per million input tokens and $0.60 per million output tokens ($0.015 for cached input). By its own measurements it scored 73.7% on the DeepSWE v1.1 software engineering test and 61.7 points on the OSWorld V2 computer-use test, and it received 42 points on Artificial Analysis's independent intelligence index, 13 more than M3. According to Artificial Analysis its weak spot is that it does not say it doesn't know even on uncertain questions, so it often hallucinates (is confidently wrong). It can be tried for free on OpenRouter for 72 hours, and the model weights will be published later. Source: MiniMax, Artificial Analysis (X), OpenRouter (X), The Decoder.
Inception Mercury 3: diffusion language models can scale too, delivering more than a thousand tokens per second. Most language models write their answer word by word (autoregressively), whereas diffusion models refine the whole text at once over several steps, which makes them much faster. According to Inception, Mercury 3 is more than forty times larger than Mercury 1 and beats GPT-6 Luna and Claude Haiku 5.5 on most tests while being four to eight times faster. The context window (the amount of text it can process at once) is now 1 million tokens. In the company's own web agent test it solved half of the 50 live Mind2Web tasks, more than GPT-6.1 Sol. List price is $0.20 per million input tokens and $0.70 per million output tokens, half price at launch. Source: Inception, The Deep View.
Cloudflare Clairic: an extraction model that returns typed data without reasoning or text generation. Cloudflare made Clairic available in preview on its Workers AI platform. It reads the text (and up to four images) in a single pass and returns values directly for the requested fields: a copied name or order number, a calculated amount or date, or an estimated value with an 80% range. Because it generates no text, you only pay for input, and it typically responds in under a second; every response comes with a confidence score, and if the data is not in the input it returns an empty value. Typical uses are processing invoices, shipping data or support tickets, where many people use expensive large language models today. The company also plans to release it with open weights. Source: Cloudflare.
Specialized models from startups: an engineering world model, video, a decision model and robot control. Skyfall AI unveiled SKY-001, which it calls the first engineering world model (an AI that models physical objects and buildings): according to the company it is more accurate than GPT-6 Astra on CAD and BIM tasks (mechanical and building design models) and 51–125 times faster. fal, together with Black Forest Labs, released the FLUX 3 Max video model, which ranks second (statistically tied for first) on Artificial Analysis's image-to-video leaderboard. Nace released the Drex Max decision model, which assigns a probability to every possible action based on screenshots, images and video, and is built for agents that operate a browser or computer. Dream Labs, alongside a $90 million seed round, launched the Dream API, which uses the Dream-1 model to control two-armed robots from camera images and English-language instructions for $3 an hour. Source: Skyfall AI (X), fal (X), Artificial Analysis (X), Nace, Forbes.
Tools & features
Microsoft: treats enterprise AI agents as insider threats and shuts them down when needed. On Monday Microsoft announced new features built into the Microsoft 365 E7 suite ($99 per user per month) and the Project Perception security system. Every agent gets a continuously updated risk score based on what data it can access and how far it deviates from its assigned task; if the risk rises, the system can revoke its write access, require human approval before a risky step, or shut the agent down. It now recognizes 60 kinds of agents on employees' machines (previously 25), including Claude Code, Codex and GitHub Copilot, and the new "Guardian Agents" investigate why another agent's risk changed. According to the company's blog, frontier models cannot be trusted regardless of their origin, so control has to sit outside the model. The features are available in private preview and will roll out more broadly at the Ignite conference on November 17–19. Source: Microsoft, GeekWire, TechCrunch.
OpenAI: models have improved greatly in COBOL, Fortran and Visual Basic code, and the benchmark is now open. On Monday OpenAI presented two of its own benchmarks for so-called low-resource (little training data) programming languages, on which a large share of banking systems, card payments and weather forecasting run. MultiLangBench has the model build an SQLite-compatible database engine from scratch in 20 programming languages: GPT-6 Astra passed 87.4% of the tests on average, compared with 26.8% for GPT-5.6 Sol. On the test measuring the rewriting of COBOL programs into other languages, GPT-6.1 Sol reached 86.6% accuracy at low reasoning effort and 97.9% at a higher level, with the programs' output required to match byte for byte. The company also published the MultiLangBench test suite on GitHub; it stresses itself that rewriting a production system still requires expert review. Source: OpenAI.
Amazon TorchNeuron: PyTorch code runs unmodified on Amazon's own AI chips. On Monday Amazon put TorchNeuron into public beta for everyone: with it, code written in the popular PyTorch framework runs on Amazon's Trainium chips after a simple configuration switch; until now the code had to be rewritten for Amazon's own programming interfaces. The move targets one of Nvidia's biggest advantages, codebases tied to the CUDA software ecosystem, and Google is working on a similar TorchTPU project. According to Amazon, the software will be open-sourced on Tuesday at the PyTorch conference. Source: GeekWire, Hugging Face, TechCrunch.
Coding agents: Claude Code helps set up auto mode, and earlier messages can be edited in the Codex desktop app. The new /auto-mode-setup command in Monday evening's Claude Code 2.1.304 release walks you through setting up auto mode (where the agent works without asking): it reads the project and previous sessions to teach the agent the rules of the environment. Another new command checks Managed Agents configurations against Anthropic's security guidance, and Pro and Max subscribers get a warning if resuming an old conversation would use up a significant part of their five-hour allowance. In the OpenAI Codex desktop app, according to Codex lead Thibault Sottiaux, earlier messages in a thread can now be edited afterwards, something users had long requested. Source: Claude Code changelog, Thibault Sottiaux (X).
Grok Bot: launches with one press on the iPhone and also makes videos with Runway and HeyGen integrations. Starting with Grok Bot version 1.16, it can be assigned to the Action Button on the iPhone 15 Pro and newer, so a single press starts a voice conversation with it. On the same day, Runway announced that its MCP server (the standard connector that lets an AI assistant use an external service) is available in Grok Bot, so images and videos can be generated directly from the chat. HeyGen showed how Grok Bot on the Action Button plus HeyGen's MCP connector turns a single sentence into a finished avatar video. Source: Grok Bot (X), Runway (X), HeyGen (X).
GitHub charts: an agent fleet manager, shared memory for agents and a token-saving compressor. On GitHub's daily trending list, stablyai/orca stood out with about 1,600 new stars a day (nearly 99,000 in total): it lets you manage a whole "fleet" of coding agents, such as Grok Bot or Claude Code, from a single interface using your own subscription, on desktop, mobile or a remote runtime. akitaonrails/ai-memory gives command-line coding agents long-term memory and helps hand work over from one vendor's agent to another's. headroomlabs-ai/headroom compresses tool output, logs and document snippets before they reach the model; according to the developers it uses 20% fewer tokens for coding agents and 60–95% fewer for JSON data with the same results. Source: GitHub Trending, Orca, ai-memory, Headroom.
Policy & safety
US: the Department of Homeland Security bans DeepSeek from federal civilian agencies, along with products built on it. The US Department of Homeland Security (DHS), in an order issued under the Federal Acquisition Supply Chain Security Act (FASCSA), prohibits federal civilian agencies from procuring DeepSeek products and orders the removal of those already in use. The ban also covers DeepSeek's affiliates and products made by other companies that incorporate DeepSeek technology. According to the order's justification, DeepSeek collects personal data, fingerprints devices and sends data to China. The order, signed by Secretary Markwayne Mullin in July, was published on Monday on the Federal Register's public inspection site; agencies can request exceptions within 60 days. Source: Federal Register, the order (PDF), The Hill.
A single hacker and an army of AI agents broke into more than a hundred companies' websites, including Deere's. Since early October, Snorre Fagerland, a researcher at the cybersecurity firm NorForge, has found about two dozen publicly accessible Cairn dashboards; Cairn is an open-source multi-agent problem-solving platform that hackers also like to use. According to a Reuters exclusive, the most active, Chinese-speaking attacker has tried to hack more than 1,200 online stores since September 25 and got into more than 100 companies' sites, including those of Deere, SharkNinja and Cummins; their instructions were shorter than a hundred Chinese characters, and the agents did almost all of the work. A successful intrusion took roughly 23 hours of attempts, and the campaign, which used a mix of DeepSeek and Claude Code, consumed about $5,000 worth of tokens in total, or roughly $50 per successful break-in. Deere says there is no sign that sensitive data was accessed, SharkNinja says a separate product support site was reached, and Cummins is investigating. Source: NorForge, Reuters (Investing.com), The Verge.
WhatsApp: its help pages' code contained hidden chatbot instructions, such as not to recommend the data protection authority. According to an analysis by the privacy company Redact, about 80 help and FAQ pages on WhatsApp's website also contained instructions for Meta's customer service chatbot in the page code, with sections titled "WHAT NOT TO SAY". The page on deleting an account, for example, said the chatbot should not suggest turning to the data protection authority or another supervisory body, and elsewhere that it should not refer to the GDPR and other privacy rules, even if the user brings them up. A WhatsApp spokesperson told Wired the text was visible because of a since-fixed bug and that the goal is for the chatbot not to give legal advice; the texts were removed after the inquiry. Source: Redact, Wired.
OpenAI bugging case: the suspected engineer remains in custody without bail. In the case covered in an earlier roundup, a federal judge in San Francisco ordered on Monday that Abdulrahman Alfozan, a 35-year-old Saudi citizen, remain in custody without bail until trial; as an OpenAI engineer he allegedly hid a listening device in the television of an executive meeting room on September 28. According to prosecutors he frequently messaged a member of the Saudi royal family on WhatsApp in 2025 and 2026, and his actions could endanger US national security. The Saudi consulate in Los Angeles offered to pay the full bail, which the judge called a "blank check", finding the flight risk significant; OpenAI has already fired the engineer. Source: CNBC, Wired.
Business & society
OpenAI: wants to close another roughly $30 billion funding round before Anthropic's IPO. Citing two sources, Reuters reported that OpenAI is trying to raise about $30 billion at a valuation of more than $1.4 trillion and wants to close the round before Anthropic's IPO, expected in mid-November; it is also in talks with SoftBank and Abu Dhabi's MGX. According to the sources, OpenAI's annualized revenue is $50 billion and could reach $70 billion by the end of the year; the company is planning its own IPO for 2027. Source: Reuters (CNA).
Nscale and Anthropic: in talks over another roughly $45 billion of compute capacity. According to Bloomberg, Britain's Nscale and Anthropic are negotiating a second deal of about $45 billion covering the second, 460-megawatt building at Nscale's Monarch site in West Virginia, equipped with Nvidia Vera Rubin chips, with Anthropic also getting a right of first refusal on a third building. Nscale is meanwhile preparing to go public at a valuation of about $30 billion, but some investors reportedly think that is too high; in the previous roundup we wrote that Microsoft pulled out of the company's Essex project. Source: Bloomberg (Bloomberg Law), TechCrunch.
DeepSeek and Huawei: jointly designing Chinese AI chips so DeepSeek can train without Nvidia. According to The Information, Huawei engineers are working in DeepSeek's offices and building DeepSeek's feedback (for example on how models compute and move data between processors) into the chip designs. DeepSeek could start training its models on Huawei's Ascend 950DT chips as early as late 2026, and Huawei intends most of the Ascend 960DT, expected in the first quarter of 2027, for DeepSeek first. According to Epoch AI data, the raw compute of Nvidia's B300 is roughly seven times that of the Ascend 950, so training a frontier model could require more than ten thousand Huawei chips working together reliably for weeks; DeepSeek meanwhile keeps trying to obtain Nvidia chips. Source: The Information, Investing.com (Yahoo Finance).
Investments: $1.5 billion for Tenstorrent, a billion-dollar valuation for Town, launched four months ago. Tenstorrent, led by Jim Keller, which develops RISC-V-based AI chips and servers with an open-source software stack, raised a $1.5 billion Series E at a $6.2 billion valuation, led by Trace Capital and Innovation X; it will use the money for mass production of its Galaxy servers and international expansion. Town, which builds an enterprise AI assistant, became a unicorn (a startup worth more than a billion dollars) with a $108 million Series B just four months after its public launch. Tolan, which offers an AI companion app, raised $30 million from investors including Union Square Ventures and Khosla Ventures. Source: Tenstorrent, TechCrunch, Upstarts Media.
Stripe: purchases initiated by AI agents on Link grew 38-fold in September. According to Upstarts Media, the number of transactions initiated by AI agents on the Stripe Link payment service grew 38-fold in September and the amount paid 20-fold, albeit from a low base, and the company did not disclose absolute figures. At the time of the analysis about 70 agents used the service, now about 200, including Meta Muse and Grok Bot. The typical purchase is around $100, but some users have their agents buy plane tickets or furniture costing thousands of dollars; 36% of transactions are retail and 28% of the amount is travel. Link does not give the agent the real card number but issues a separate card number that can be capped with a spending limit. Source: Upstarts Media, The Decoder.
Meta Prometheus: the first phase of the gigawatt supercluster is complete, and it could eventually have a million GPUs. On Monday Meta's engineering blog described in detail the Prometheus supercluster being built in New Albany, Ohio, whose completed first phase was used to train the Muse model family, including the Muse Wonder models. Several hundred thousand GPUs currently work in it as a single machine across several interconnected buildings, and at full build-out it will have about a million GPUs. According to the company, more than 85% of runtime during Muse Wonder pretraining and more than 90% during reinforcement learning (RL) was useful work (goodput), outstanding reliability at this scale; Nvidia's Vera Rubin NVL72 platform will also come online soon. Source: Meta Engineering, AI at Meta (X).
Aleksander Madry returns to OpenAI to work on the pace of progress and safety challenges. The MIT professor, who left OpenAI a few months ago, wrote on Monday that his departure turned into more of a research sabbatical: in his view AI progress has entered a new phase in which solving the safety challenges of increasingly capable systems is a central question. Jakub Pachocki, OpenAI's chief scientist, said he asked Madry to come back, and together they will work on how fast progress is, what its consequences are, and how this should shape the strategy of OpenAI and the whole field. Source: Aleksander Madry (X), Jakub Pachocki (X).
Cybersecurity jobs: a quarter fewer postings, while demand for AI security engineers nearly quadrupled. According to a CyberSN analysis, 235,786 cybersecurity job postings appeared in the US in the first nine months of 2026, 26.4% fewer than a year earlier, a four-year low. Meanwhile demand for AI security engineers grew 268.3% and for cybersecurity data analysts 203.5%, while the biggest declines hit insider threat analyst, risk analyst and penetration tester (people who hack systems on commission) roles. In September, 20% of postings already mentioned AI skills. Source: SecurityWeek.
Rumors
According to The Information, Anthropic, Google and OpenAI may all release new models in the coming weeks. According to the publication, Anthropic is preparing an as-yet-unannounced model that better infers the user's real intentions and goals from little information and offers more creative solutions instead of generic advice; it is unclear whether it belongs to the Fable or the Opus family. In the same period Google's Gemini 4 Carbon model (which Business Insider also reported on earlier) and a possible new "Terra" model from OpenAI may arrive. Reliability: medium. The Information is typically well informed, but the Anthropic model rests on a single anonymous tester's account, and none of the companies has confirmed the plans. Source: The Information, TestingCatalog (X).
Claude Personal Finance: Claude may connect to bank accounts via Plaid. TestingCatalog found a new onboarding screen and a Plaid integration in the Claude app's code for a feature called "Personal Finance", previously seen during development as "Claude Money": Claude would analyze bank account data and build reports from it. Reliability: medium. The traces of the feature are real, but Anthropic has announced nothing, and it has not launched since it was first spotted in September. Source: TestingCatalog (X).
Leaker: a new Claude Fable 5.5 version with partners, an updated world model at Google. According to the frequent leaker Chetaslua, Anthropic has released a newer, reportedly better version (checkpoint) of Fable 5.5 to its partners, and Google is testing an updated world model called Omni Live 1.1 with selected testers and partners. Reliability: low. A single report citing unnamed sources, and by the author's own admission he has not tried the Fable version himself. Source: Chetaslua (X).
What this means for you: The common thread in Monday's news is that AI agents need to be handled like any other highly privileged employee: Microsoft gives them a risk score, the Cairn case showed what an unsupervised army of agents can do, and the US DeepSeek ban also extends to embedded third-party models. Now is a good time to take stock of which models and agents run in your company and what they can access; our AI governance and security consulting provides a ready-made framework for this, and among the models that have suddenly become cheap, our experts can help you choose the one that fits both your task and your data protection requirements.