Mistral shipped Leanstral 1.5 on July 2. It is Apache-2.0-licensed, free during the beta, and built to generate formal proofs in Lean 4, the proof assistant that mathematicians use to check theorems and that a small but growing number of software teams use to check that their code actually does what the spec says it does. The model is 119 billion total parameters with a mixture-of-experts architecture that keeps only about 6 billion active per token, which is what lets Mistral charge nothing during beta and still plausibly plan to charge something later that is not embarrassing.
The headline number is that Leanstral 1.5 beats Claude Sonnet by 8 points on FLTEval, the proof-engineering benchmark Mistral shipped alongside the original March release, at pass@16, while costing roughly 15x less to run the same query. That framing is a little Mistral-flattering (FLTEval is a Mistral-built eval, and pass@16 is a generous sampling budget), but the underlying claim is holding up in independent testing. A model that can spend sixteen tries on a Lean proof and come back with a valid one 8 percentage points more often than a Claude of similar generation, at a fifteenth of the cost, is a real product.
The line in the release that is worth stopping on is the one where Mistral notes, in the polite present tense, that the model “has uncovered previously unknown bugs in open-source code.” Nobody is quite naming the codebases. Nobody is quite quantifying the count. But the mechanism is straightforward and slightly awful: point a Lean-fluent agent at a widely-used open-source library, ask it to prove that the library actually behaves the way its documentation says, and watch it produce counterexamples. Those counterexamples are bugs. Some of the libraries whose proofs failed are libraries that a lot of people are currently running in production, and no one has told them yet.
This is the correct product to build. It is also the moment where the “AI writes code” conversation starts to invert. Most of the discourse for the last three years has been about whether AI-generated code is good enough to trust. Leanstral 1.5 is a bet that the more interesting question is whether human-generated code was ever good enough to trust in the first place, and that a machine that can do formal verification cheaply enough to run continuously against every dependency in your build tree is not a coding-assistant story. It is a supply-chain-security story wearing coding-assistant clothes.
The near-term implication for anyone actually shipping software is straightforward and uncomfortable. Somewhere in the next twelve months, a security researcher, or a bored graduate student, or Mistral itself, is going to publish a batch of previously-unknown vulnerabilities in a popular OSS library, disclose them responsibly, and cite Leanstral (or whatever the next one is) as the tool that surfaced them. And then every CISO is going to have to answer a question about whether their org runs continuous formal verification on its dependencies. And the honest answer, for approximately all of them, is going to be no.
Free during beta. Apache 2.0. The bill comes later, one way or the other.