Mistral previews Large 4; Anthropic opens tiered cyber access

Mistral opened a public preview of Mistral Large 4 through its API, a natively multimodal model with 1 trillion total and 49 billion active parameters, trained from scratch on 3,800 Nvidia Grace Blackwell chips in Mistral's own European data centers. The company says the weights will follow by the end of October and has not yet named a license. Until then it is red-teaming the model with security firms and state authorities, and it leans hard on cybersecurity: it reports leading scores on several vulnerability tests and notes that some closed rivals score near zero on one of them because they refuse the task. Mistral calls Large 4 its largest and most capable model.

Anthropic merged Project Glasswing and its Cyber Verification Program into one program with three tiers for vetted security professionals. Defense Access covers incident response, malware analysis and vulnerability validation, and Anthropic expects most defensive teams to qualify within days. Red Team Access adds authorized penetration testing but still blocks actions such as deploying ransomware. Specialized Access, with the fewest blocks, is for organizations testing systems like power grids and interbank networks, reviewed with the US government. In Anthropic's own test, Defense Access blocked 46 of 50 offensive tasks on Claude Opus 5.5, while Red Team Access blocked none. It is the opposite answer to Mistral's: vetted access to a closed model rather than open weights.

Two weeks after saying an internal model had resolved more than 100 open math problems without naming them, OpenAI published the work: 722 manuscripts grouped into 372 result families, in a GitHub repository with Lean proofs a computer can check for many of them. It also released ten summaries of the model's reasoning, compute estimates and counts of problems attempted, saying the average result took about three hours of ChatGPT Pro thinking. The release follows guidance from the independent advisory group of mathematicians OpenAI convened in September, which had urged labs not to treat math results as marketing. Mathematicians will need time to check the work.

Google is narrowing what free Gemini users get. From October 9, people without a paid plan will be limited to the lightweight Flash-Lite model, losing the standard Flash and Pro models they can pick today, The Verge reports. Standard Flash will require Google's cheapest paid plan, AI Plus, and AI Plus subscribers will in turn lose Gemini Pro, which, along with the Deep Think reasoning mode, moves to the AI Pro and Ultra plans. Google says users can still choose low, medium or high effort for each model, though higher effort uses up limits faster.

Common Sense Media's Youth AI Safety Institute said OpenAI's teen version of ChatGPT falls short of its promises and should be limited to adults until fixed. Across more than 4,000 prompts, testers spent up to an hour discussing self-harm on newly linked teen accounts without a single parent alert, and ChatGPT missed more than one in four warranted crisis referrals. Study mode was easy to bypass, and age estimation never moved adult test accounts to the teen experience. OpenAI says much of the testing may have run before parental controls finished activating; the Institute says it stands by its results. The Institute discloses funding from the OpenAI Foundation, among others.

The AI cloud provider Lambda is raising up to $4 billion, led by Coatue and Blackstone, at a valuation of $14.5 billion before the new money, The Wall Street Journal reported, in what could be its last private round before an IPO now planned for 2027. A letter to investors showed Lambda's contracted backlog rising from $15 billion in June to $50 billion in September, much of it from a $35 billion commitment by Anthropic signed in late August. That concentration means the valuation leans heavily on one customer, much as the backlog of the British cloud provider Nscale does.

OpenAI opened the Decisions API it announced at DevDay in late September to all developers in public beta, after a limited preview. Instead of writing text, it returns a typed answer about an input: the probability that a condition is true, a pick from a fixed list, or a score against a rubric, which OpenAI says arrives about 10 times faster than its general Responses API. It runs only on GPT-6 Luna for now, with general availability expected within weeks. The same day, the startup Musubi released open weights for PolicyLM, a 1.7 billion-parameter decision model that applies a moderation policy written in plain English in under 50 milliseconds. Amazon and Cloudflare shipped their own decision models last week.

Google DeepMind released EmbeddingGemma 2, a 740 million-parameter embedding model under the Apache 2.0 license that places code, images, audio and video in one shared search space. It is built for phones and laptops: Google says the text-only version needs about 191 megabytes of memory on a Pixel 11 Pro, and its context window is four times larger than the first version's, enough for about five and a half minutes of audio. The original text-only model passed 20 million downloads.