Hope your weekend's going well — here's your weekly AI rundown 👋

This week AI did something mathematicians thought would take years, and did it in eleven days. Meanwhile Europe wrote its biggest-ever tech check, and everyone who has to actually buy this stuff hit a wall the industry is now calling "model fatigue." Big leaps, big money, and a growing sense that the pace is the story.

Andrew Wiles proved Fermat's Last Theorem back in 1994 — but translating that proof into a fully machine-checked formal proof (written in the Lean programming language, so a computer can verify every step with zero hand-waving) was a project experts figured would take years. On September 5, Anthropic reported that Claude produced the first fully machine-checked proof in just 11 days: about 13 million lines of Lean, 29,500 intermediate theorems proved along the way, dozens of Claude agents working in parallel, and roughly 6 billion tokens. It's the largest formal proof ever assembled. Why it matters beyond the flex: formal verification is math you can trust absolutely — and the same technique that checks a theorem can one day check that safety-critical code does exactly what it claims.

France's Mistral closed a €3 billion Series D led by Samsung, valuing it above €21 billion (~$24B) — the largest equity fundraise in European tech history. NVIDIA, ASML, BlackRock and even the government of Luxembourg joined. CEO Arthur Mensch says the money goes toward building and owning data centers. "Sovereign AI" — the idea that regions want frontier models they control — just became a very big business.

Our take: the same week a model did years of math in eleven days, Europe wrote its biggest-ever check to make sure the next frontier model isn't only American. Capability and capital both took a step up — and neither is waiting for the other to catch its breath.

The AI called him a criminal — and a court is listening. Activist Robby Starbuck won a preliminary ruling against Google on September 8 after its AI allegedly described him as a child molester. The case is one of several forcing courts to wrestle with an awkward question: defamation law hinges on intent and a "state of mind," and a language model has neither. Who's liable when the machine makes something up about a real person? Nobody has a clean answer yet — and the answers are starting to come from judges, not engineers.

The week in five figures:

With a new frontier model landing roughly every 11 days, the hard part isn't finding one — it's choosing. This week's collection cuts through the noise:

→ Best LLMs (2026): the 8 models powering AI today — GPT, Claude, Gemini, Mistral and more, side by side: what each is genuinely best at, where it costs the most, and which one to reach for by task. If model fatigue is setting in, start here.

You know ChatGPT and Claude. Here are five lesser-known tools worth adding to your kit:

This week's spotlight: Mistral AI. Fresh off the biggest funding round in European tech history, the French lab is the clearest sign that frontier AI won't be a US-only story. Its pitch isn't just another chatbot — it's Vibe, an agent platform built for long-horizon work like research, drafting and writing code, running on models Mistral trains itself and, increasingly, on data centers it owns. Open weights, European data residency, and now €3B to scale all of it. Watch their short look at Vibe below.

New arrivals in the AI Graveyard:

Want the autopsy? We dug into what actually kills AI tools — lost funding, acquisition, or just a dead domain — in our data report.

When the machines outrun the manuals

Four frontier models shipped inside one 72-hour window this month, and the median gap between major releases has compressed from about 37.5 days in 2023 to roughly 11 today. For the people who buy and deploy this stuff, that's the "model fatigue" everyone's suddenly naming — you barely finish testing one model before it's deprecated for the next. Sam Altman's own summary: "we're all moving to faster cadences."

The quieter cost isn't inconvenience, it's safety. Documentation, red-teaming and integration testing all get squeezed into that shrinking window — in the very same weeks AI is formalizing century-old proofs and finding zero-day exploits. More than 1,000 lab employees have now signed a petition asking their own employers to ease off the gas.

Speed has become the product. The part that can't be rushed — making sure the thing is safe and understood before the next one ships — is exactly the part getting compressed. Worth watching whether "faster cadences" ever collides with a mistake big enough to slow everyone down.

Can't tell which model to use for what? Paste this in and let it sort you out:

On the calendar:

Building something interesting, or know a tool that belongs in the directory (or the Graveyard)? Hit reply, or email [email protected] — we read everything. (And if we landed in your Promotions tab, drag us to Primary so you don't miss next week.)

— The ToolDirectory.AI team