Mistral AI's Leanstral 1.5 masters math proofs with big cost savings

European AI company Mistral AI has launched Leanstral 1.5, a model designed for formal math proofs. It has 119B total parameters, activates 6B, and is open-source under Apache-2.0.

The model aced math tests: 100% completion on miniF2F, solved 587 out of 672 PutnamBench problems, and hit 34% on PhD-level FATE-X—all best scores.

Mistral AI highlighted its cost edge: Leanstral 1.5 averages $4 per problem on PutnamBench, compared to over $300 for Seed-Prover 1.5 and $54-68 for Aleph Prover.

In real-world use, it flagged 47 violations across 57 codebases, with 11 pointing to real bugs, including 5 previously unreported issues on GitHub.