Daily AI Brief: Mistral's Large 4, OpenAI Math Gains, and Open-Model Pricing Debate
Mistral introduced Large 4, described as a trillion-parameter multimodal model with 52 billion active parameters, trained on NVIDIA Grace Blackwell hardware in European data centers and positioned around European AI sovereignty. The roundup also covers OpenAI mathematics results, discussion of DeepSeek 4.1 Flash and Chinese open-model pricing, Andrej Karpathy's comments on differing views of LLM abilities, and new questions about whether coding agents increase software output. Mistral's open weights are expected by the end of October.
Mistral's Large 4 entered public preview on October 6 as a natively multimodal system with one trillion parameters, of which 52 billion are active. It was trained entirely on 3,800 NVIDIA Grace Blackwell GPUs in European data centers, with inference also hosted there; open weights are expected by late October. Reported strengths include coding, agentic tasks, visual grounding, and cybersecurity.
OpenAI released a set of mathematics results generated by an internal frontier model, including Lean formalizations and a proof of the Unique Games Conjecture. The company estimated an average result consumed compute comparable to about three hours of ChatGPT Pro thinking and consulted an independent mathematics advisory group at the Institute for Advanced Study.
Mistral's European-hosted model and promised open weights could give developers and regulated organizations more options for keeping AI workloads within a chosen jurisdiction. OpenAI's mathematics results may accelerate research workflows and change how some proofs are checked or taught. Meanwhile, debates over open-model pricing and coding-agent productivity could shape which businesses, researchers, and smaller teams can afford advanced tools, though actual societal effects will depend on adoption, reliability, and governance.