A free, open model now proves your code is correct! It found five real bugs nobody had reported
Mistral released Leanstral 1.5, an open-weight coding agent built for Lean 4, a proof assistant, not a chatbot. It's Apache-2.0, the weights are on Hugging Face, and there's a free API endpoint. Beyond the maths benchmarks, Mistral ran it across 57 open-source repositories: it flagged 47 violated properties, 11 of which were genuine bugs, and five of those had never been reported on GitHub. It also proved the O(log n) time complexity of a real AVL tree implementation, a run that took over 2.7 million tokens.
What it actually means
Every other story this month has been about generating more code, faster, for less. This is the counterweight, and it's the more interesting half of the thread. When your agent writes ten times more code than you can read, the bottleneck stops being generation and starts being trust and "the tests pass" is a much weaker claim than most teams pretend. A prover doesn't guess whether your code is right; it either compiles a proof or it doesn't. But if you have one function where being wrong is genuinely expensive - money maths, an auth check, a parser eating untrusted input - that's now something you can hand to a free model and get a machine-checked answer, not a vibe. Watch this direction rather than this model. The agents that get trusted with production won't be the ones that write the most; they'll be the ones that can show their work.
When the machine writes the code, "the tests pass" stops being good enough.
Source: Mistral AI · Monday, 13 July 2026