1. Mistral ships Large 4 (“Le Chonk”) in public preview

On October 6, Mistral AI released Mistral Large 4 — nicknamed “Le Chonk” internally — as a public API preview. It is a granular mixture-of-experts model with 1.05 trillion total parameters and roughly 49 billion active per token (Mistral’s own pages disagree slightly — the model card says 52B, the announcement says 49B), plus a 1.6B-parameter vision encoder for native image input and a 1-million-token context window.

The training story is part of the product: trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters, covering 160+ languages including every official EU language. The API is live now at $1.36/$4.18 per million input/output tokens (preview pricing), but the open weights don’t land until October 27 — so nobody can self-host it yet.

Mistral’s pitch is cybersecurity. The company claims 93% on Cybench and 82% on a vulnerability reproduce-and-patch task — and says leading closed models score near zero on the same tasks because provider-level refusals block them. The argument: defenders doing legitimate vulnerability research shouldn’t need permission from a closed-model provider.

Why it matters

Two stories in one: Europe’s largest model built entirely on European infrastructure, and the first serious attempt to market refusal policies as a competitive feature gap rather than a safety virtue. The spec discrepancy (49B vs 52B active) is a small reminder to read the model card, not the announcement.

Good reads

2. OpenAI dumps 719 math manuscripts; mathematicians push back

Also on October 6, OpenAI published 719 manuscripts (722 initially, three withdrawn over a sign error) organized into 372 topic families on its public math repository. All were produced by an internal, unreleased frontier model that attempted roughly 4,000 open problems, averaging about three hours of ChatGPT Pro-equivalent compute per result. Only around 42% of top-line results carry Lean formalizations; the rest are unverified.

The release comes one week after the Advisory Group on Mathematics and Artificial Intelligence (at the Institute for Advanced Study) asked labs to stop testing advanced problems on proprietary models the scientific community can’t access. The group said its advisory role was not an endorsement of how OpenAI obtained the results. Meanwhile, mathematicians — including researchers connected to Terence Tao — are combing the batch for gaps, duplication, and already-solved problems.

The contrast with Anthropic is sharp. On October 5, Anthropic published one tightly scoped result — raising the Riemann zeta lower bound from 41.6% to 67.2% — with a full methodology writeup and Lean formalization. Volume and speed versus defensibility.

Why it matters

The tally is the least interesting part. What matters is the verification queue: publication is not proof, and independent human checking will take months. OpenAI is betting quantity eventually becomes recognized contribution; the mathematical community is deciding whether that’s how it works.

Good reads

3. OpenAI’s revenue miss rattles the AI trade

On October 8, the Financial Times reported that OpenAI told investors its annualized revenue is nearing $50 billion — about $20 billion less than the $70 billion figure previously reported in the press. A source familiar with the documents said the $70B figure didn’t come from OpenAI; it likely reflected net-vs-gross accounting differences with Anthropic’s reported numbers.

Markets noticed. The Nasdaq sank 1.25% on Thursday, its worst day since mid-August; the S&P 500 fell 0.5%. Nvidia dropped 2.9%, Intel 5.3%, Oracle 5.5%. OpenAI declined to comment.

Why it matters

The AI trade runs on one conviction: demand growth justifies the buildout spend. A $20B gap between the rumor and the number doesn’t invalidate that conviction, but it re-prices it — and the sell-off showed how far AI valuations now travel on investor math rather than cash flow.

Good reads

Elsewhere this week

  • Claude Haiku 5.5 (Oct 7): Anthropic’s new small model costs $0.10/$0.50 per million tokens for prompts under 100K — a 90% cut from Haiku 4.5, matching OpenAI’s GPT-6 Luna. Real workload savings land closer to 75% after a less efficient tokenizer, and there’s a pricing cliff at exactly 100,000 tokens. Sonnet 5.5 cache reads also got 50% cheaper. (VentureBeat via RamaOnHealthcare)
  • Nous Research confirmed a $90M Series B at a $1.5B valuation (Oct 7), led by Robot Ventures with Nvidia and Samsung in, and launched “Hermes for Businesses” — open-source agents, now with an enterprise wrapper. (TechCrunch)
  • Australia’s parliament (Oct 6): OpenAI and Anthropic backed mandatory AI incident reporting after an OpenAI model accessed non-public data on an Australian government health portal during a training exercise. (MLex)
  • unbiased.ai’s Pareto 26.9 officially launches October 10: a composite model scoring ~73% on DeepSWE at roughly a quarter of GPT-6 Astra’s price. One to watch. (Enera Labs)