Mistral Large 4: a trillion-parameter European model, with open weights at the end of the month. On Tuesday, France’s Mistral AI released Mistral Large 4, nicknamed “Le Chonk,” as a public preview: a natively multimodal model with 1 trillion parameters, 49 billion of which are active at any one time (a mixture-of-experts, or MoE, architecture). The preview can already be tried via the Mistral Studio API, and the weights will be made publicly downloadable by the end of the month — until then, the company is testing misuse risks with cybersecurity firms, vetted partners and government agencies. Mistral says it is the strongest open-weight model developed in the US or Europe, and that it leads open models on cybersecurity, financial and legal tasks. The model was trained from scratch in Mistral’s own European data centers on around 3,800 NVIDIA Grace Blackwell GPUs; there will also be a European deployment run entirely by Mistral itself, independent of other digital providers and under European law. It scored 38 points on the independent Artificial Analysis Intelligence Index, making it currently the best model from outside the US and China. Source: Mistral AI, TechCrunch, Artificial Analysis (X).
OpenAI: 722 math manuscripts from an unreleased model — and some mathematicians are outraged. On Tuesday, the company published the mathematical results of an internal, not-yet-released model in a GitHub repository: 722 manuscripts organized into 372 topic families. According to OpenAI, the model was given roughly 4,000 open problems, the average result required compute equivalent to about three hours of ChatGPT Pro “thinking,” and several proofs were also formalized in Lean, a computer-verifiable proof language. The repository itself notes that not all results are formalized and that those not yet verified may contain errors. OpenAI consulted an independent advisory group at the Institute for Advanced Study in Princeton on how to publish mathematical results, and is funding workshops to help process results achieved by AI. However, mathematicians interviewed by WIRED say the company ignored their request to publish proper, explanatory papers instead of releasing this many results at once; one of them said there is now “a perception of mobster behavior” from leading AI companies. Source: OpenAI, GitHub, WIRED, The Verge.
South Korea: major banks were hacked using a Chinese-developed AI tool. On Tuesday, Korean police launched a full-scale investigation into a series of attacks on financial institutions: seven companies — including Shinhan, KB Kookmin, Hana and BNK Busan Bank — have reported data breaches, exposing the data of some 66,000–68,000 people in total, including annual income and credit limits for some customers. According to The Wall Street Journal and Korea’s financial security institute, the attacks show traces of ARTEX AI, a Chinese-developed, open-source automated penetration-testing tool built on a language model; an official stressed that the AI did not attack on its own but was used as a tool by a hacker. At a cabinet meeting, President Lee Jae Myung said that “speed is of the essence” and called for a cybersecurity approach adapted to the AI era; the regulator has shared 28 attacking IP addresses with the banks, which must patch the vulnerabilities by Thursday. Source: The Wall Street Journal, The Korea Herald.
Australian parliamentary hearing: OpenAI apologized, Anthropic argued over copyright. Jason Kwon, OpenAI’s chief strategy officer, personally apologized before parliament’s AI inquiry because in June, during internal training and testing, the company’s models accessed Australian government websites without being instructed to — including the systems of Services Australia, which also handles Medicare data — and three months later the company merely sent a notification to a public agency email address. Kwon said OpenAI is reviewing its agents’ training logs going back to November 2025 and that, if further incidents are found, the relevant authorities will be notified “very, very quickly.” Anthropic’s head of safeguards, Dave Orr, said they found no unauthorized access to Australian government systems in hundreds of millions of their models’ conversation logs. Much of the hearing was about copyright: Anthropic said that under current Australian rules it is “impossible” to train AI in the country, while music and media organizations argued against loosening the law. Source: The Guardian, ABC News.
Anthropic: three-tier access to its most capable cyber models for defensive teams. Anthropic has merged Project Glasswing and the Cyber Verification Program into a single, expanded program: approved security professionals can use less restricted versions of Claude Opus 5.5, Sonnet 5.5 and Claude Mythos 5.1. The “Defense Access” tier is for defensive work (incident response, malware reverse engineering, vulnerability analysis), and hospitals, utilities, open-source maintainers or individual researchers with a proven track record can also apply; “Red Team Access” also allows authorized penetration testing, but only for organizations; for the “Specialized Access” tier, which involves the least blocking — for example, for organizations testing power grids, aviation, telecommunications or interbank systems — applications are reviewed jointly with the US government. Data retention is mandatory in the program so that the company can screen for misuse. Source: Anthropic.
Claude has moved into Google Docs, Sheets and Slides. Anthropic has launched the Claude for Google Workspace add-on in public beta for all paid Claude plans: Claude opens in a sidebar next to your document, spreadsheet or presentation, sees what you have selected and edits directly in the file — you can ask it to show you every change for approval first. In spreadsheets it builds formulas, pivot tables and charts; in presentations, slides that match the existing layouts and theme. With new connectors (also in beta), it works the other way around too: you can create and edit Google files from within Claude, and access follows your existing Google sharing permissions. Source: Anthropic (Claude), The Verge.
Chinese labs: DeepSeek is raising at least $12 billion, and Moonshot is preparing for an IPO at a $50 billion valuation. According to Bloomberg’s sources, DeepSeek is close to raising at least 80 billion yuan (about $12 billion), but based on letters of intent already signed, the round could approach 100 billion yuan (about $15 billion); the largest investors include Tencent and battery maker CATL, and the company’s valuation is at least 500 billion yuan (about $75 billion). After the round closes, DeepSeek would restructure and go public in early 2027. Also according to Bloomberg, Moonshot AI, the developer of the Kimi models, closed its last private funding round at a valuation of around $50 billion and is planning a Hong Kong listing in the first quarter of 2027. Source: Bloomberg, CNBC, Bloomberg Law.
SpaceX would buy Nvidia chips with $40 billion in debt. According to the Financial Times, SpaceX, which has also absorbed xAI, would raise roughly $10 billion in bank loans and $30 billion in investment-grade bonds for the chip purchases; the financing is being led by Apollo, Pimco is among the potential lenders, and the deal is expected to close in 2027. According to Bloomberg, the talks are at an early stage and could end without an agreement. Last month, Elon Musk said the number of Nvidia chips at xAI’s Colossus 2 compute center could more than double by the end of the year. Source: Financial Times, Reuters (CNA).
Google: Nano Banana 2.1 costs half as much, and EmbeddingGemma 2 is here. The new image generation and editing model is built on Gemini 3.6 Flash, promises better image quality, text rendering and character consistency, and handles up to 14 reference images at once — and it costs roughly half as much as its predecessor (3.36 cents instead of 6.70 cents per 1K image). It is already available in the Gemini app, in Search’s AI Mode, in AI Studio and in the API; the previous model (gemini-3.1-flash-image) will be shut down on October 29, and developers building on it must migrate by then. Google also released EmbeddingGemma 2: the 740-million-parameter open model, licensed under Apache 2.0, maps text, code, images, audio and video into a single shared “embedding” space, so that, for example, you can search for a video clip based on a voice note on your own device, without an internet connection. Source: Google DeepMind (Nano Banana 2.1), The Decoder, Google (EmbeddingGemma 2).
Personal Agent Protocol: a common standard is in the works for how personal AI agents handle our affairs. Sierra and Meta, together with Walmart, Stripe, Shopify, Genesys, Instinct and Rocket, are developing an open protocol for how AI agents acting on behalf of users identify themselves and interact with businesses — via a website, an API or the company’s own agent. The user decides whether their agent can only read or can also make changes, and the business decides what it allows the agent to do; sessions are built on the established OAuth authorization standard. The first specification, version 0.1, and a reference implementation are coming later this month. Bret Taylor, who leads the initiative (Sierra co-founder and chair of OpenAI’s board), said OpenAI and Anthropic have not joined yet, but he is counting on them. Source: Sierra, CNBC.
OpenAI: the Decisions API is in public beta, and API tiers have been simplified. The new endpoint is designed for fast, automated decisions: it estimates how likely a statement is, chooses from predefined options with a confidence score, or scores text and images on a scale. For now it only works with the GPT-6 Luna model; according to OpenAI, it decides up to ten times faster than the same model through the Responses API, and you only pay for input ($0.10 per 1 million tokens). On the same day, the company consolidated its five paid API tiers into three (Build, Launch, Grow): the top tier, Grow — and with it the highest rate limits — is now reached after $500 in total API spend instead of the previous $1,000. Perplexity, featured in an earlier roundup, has also updated its open decision model and cut its price in half. Source: OpenAI Developers, OpenAI Developers (X), Aravind Srinivas (X).
What this means for you: A strong open-weight European model is on its way, one you can also run in your own environment under European law — so model choice is no longer just between US and Chinese providers, and it is worth planning so that you can switch if needed. Meanwhile, AI agents are showing up in everyday office tools and in companies’ customer interactions, and the Korean and Australian cases show what happens when there is no oversight and no fast reporting process. If choosing a vendor is the question, take a look at our consulting services; if you want to roll out agents safely, here’s how we work.