Mistral put Large 4 on a public API yesterday: a trillion-parameter mixture of experts, 49 billion of them active, pictures in and text out, trained on 3,800 Grace Blackwell GPUs in the lab's own European halls. Anthropic opened three doors in the wall around Claude's hacking skills. OpenAI emptied a folder of math onto GitHub from an internal model, and linked last month's still-disputed Navier-Stokes claim.
In the news
Mistral Large 4: API this week, weights end of month
Mistral opened a public preview of Large 4 yesterday. The internal name is Le Chonk: one trillion total parameters, 49 billion active, natively multimodal, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters, and served on that same stack.1 Weights come by the end of the month, after a red-team with cybersecurity leaders, vetted partners, and state authorities. You cannot pull them this morning.
Artificial Analysis scores Intelligence Index 38, the independent number, level with GPT-6 Luna at its max, and a Cyber Index of 50. Coding and cyber charts on Mistral's own post are the company's. API list price on AA is $1.36 in and $4.18 out per million tokens, half off for the first two weeks.2 The labs have not agreed on the context window.
Anthropic Cyber Verification Program: three defender tiers
Anthropic merged Project Glasswing into an expanded Cyber Verification Program yesterday, with three access tiers that loosen Claude's cyber classifiers for defenders. Defense Access is for SOC work, incident response, malware reverse engineering, and vulnerability validation on systems you own. Individuals with a reporting track record can apply, and review is a few days.
Red Team Access adds authorized pentest for organizations only: review is a few weeks, applicants sit in Defense while they wait, and ransomware, physical damage, and high-risk safety-system tests stay blocked. Specialized Access is the fewest blocks, with US-government co-review, for flight ops, grids, telecom, interbank, and government admin networks.3
The program covers Claude Opus 5.5, Sonnet 5.5, Mythos 5.1, and future models on Claude Platform, Vertex AI, and Microsoft Foundry. Existing Glasswing members keep current-model access without reapproval. Generally-available Claude still blocks most of this, and data retention is required until Enterprise Frontier Safeguards later this fall. The application form is live.4
Anthropic's own CyScenarioBench on Opus 5.5 says the public model blocked every task on the first prompt, Defense let 4 of 50 through, and Red Team completed 34 of 50 with no blocks. Glasswing partners reported at least 129,000 verified vulnerabilities from April through July. Those counts and that bench are the company's.
OpenAI mathematics dump: GitHub corpus from an internal model
OpenAI published a research post yesterday plus a GitHub repo of mathematical results from an internal frontier model. After consulting the IAS Advisory Group on Mathematics and Artificial Intelligence, the lab put Lean formalizations, reasoning summaries, and ChatGPT Pro-equivalent compute notes in the repo.5 Average result, they say, is about three hours of ChatGPT Pro thinking. They will fund workshops and are working to release the model later; this is a corpus, not a new chatbot.6
The post links September's Navier-Stokes page. That earlier claim is still disputed, and this GitHub dump does not independently confirm it.
