Mistral Drops Leanstral 1.5 for Math-Verified Code

The software industry is actively moving away from raw AI code generation toward automated, bulletproof code verification. On June 30, European AI powerhouse Mistral officially launched Leanstral 1.5, an open-source AI model designed explicitly for formal software verification . Moving beyond standard text-based predictive coding, the model integrates with the Lean 4 programming language to provide mathematical proof that a piece of software will behave exactly as intended under critical enterprise loads . Leanstral 1.5 is available for free under an Apache-2.0 license, with 119 billion total parameters and only 6.5 billion activated during processing, making it both performant and cost-effective .

The performance leap is substantial. The model saturates the miniF2F formal mathematics benchmark, solving 587 out of 672 PutnamBench problems, and achieves state-of-the-art results on FATE-H (87%) and FATE-X (34%) . In terms of cost efficiency, Leanstral 1.5 averages just $4 per problem on PutnamBench—compared to over $300 for ByteDance's Seed-Prover 1.5—making large-scale formal verification economically viable for the first time . Beyond pure mathematics, the model excels in real-world code verification: across 57 open-source Rust repositories tested, Leanstral 1.5 uncovered 47 suspicious properties, 11 of which were confirmed real bugs, with 5 previously unreported on GitHub . The model's multi-turn environment allows it to receive a theorem statement, submit a proof, receive Lean compiler feedback, and refine its approach until the proof compiles or the budget is exhausted . In code agent mode, it works like a developer in a raw filesystem, editing files, running bash commands, and using the Lean language server to inspect goals and errors in real time .

Traditional manual debugging is becoming obsolete. In daily software engineering environments, developers are using Leanstral to instantly verify background code logic—drastically lowering cyber-vulnerability risks and ensuring high reliability for high-stakes software infrastructure before it goes live. The model's ability to convert human-readable mathematical explanations and software specifications into Lean 4's formal language bridges the gap between natural language and machine-verifiable proof, extending formal verification beyond specialists to practical engineering work . July 2026 marks a pivotal shift in AI-assisted software development. The era of purely generative code is ending. The era of mathematically verified, bug-proof AI-generated code is already here.