Mistral Large 4 has a stated size, pricing, and weight-release timing, and Anthropic's cyber program now has three access tiers. Both change which models and capabilities engineers can use.
Key takeaways · AI-distilled
Mistral Large 4 runs only 49 billion active parameters out of its 1 trillion total and was trained on 3,800 NVIDIA Blackwell GPUs in Mistral's European data centers. Weights are held back until later this month for more red-teaming.
Mistral claims Large 4 is competitive with the strongest open models and leads open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → models on several cybersecurity and enterprise benchmarks, aiming it at coding, agentic and cyber workloads.
Approved members of Anthropic's expanded Cyber Verification Program get Claude Opus 5.5 and Mythos 5.1 with fewer cyber safety blocks, with how much is relaxed depending on their verification level.
DeepSeek is reportedly close to raising $12 billion in a Tencent-led round, possibly $15 billion at a $75 billion valuation, to fund models and a data center with 160,000 Huawei accelerators ahead of a planned early 2027 IPO.
The newsletter notes OpenAI published a GitHub repo of over 700 open math problems it says its models solved, including claimed progress on the Riemann hypothesis and Hodge conjecture and a faster matrix multiplication algorithm.
Terms in this piece · Glossary
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
