Leanstral 1.5 in one sentence: an AI proof engineer, not a toy model
Leanstral 1.5 is an open-source AI code agent, built specifically for Lean 4 proof engineering, that writes and fixes machine-checkable proofs as if it were a developer working inside your repository rather than a passive autocomplete tool. Mistral AI has released Leanstral 1.5 as a new open-source code agent model for Lean 4 formal proof engineering, automated theorem proving, and autoformalization, exposing it via the labs-leanstral-1-5 endpoint and downloadable weights. The model carries 119B total parameters with 6.5B active per token, a 256k context window, multimodal input, and text output, and it is priced at USD 0 (approx. RM0) in Labs during its experimental run. This is not a generic chat model dressed up with a Lean prompt; it has been trained in interactive environments that mirror real proof work, including compiler feedback loops and filesystem-level actions inside Lean projects.
Why this release matters for Lean 4 proof engineering and formal verification tools
Lean 4 is a proof assistant and programming environment where every theorem and program property must be justified through machine-checkable proofs, which makes it powerful for verified software but punishingly meticulous for humans. Most proof engineering time is not spent on deep ideas but on translating intuition into Lean syntax, filling in missing lemmas, and chasing compiler errors; AI prover systems such as the Leanstral 1.5 model aim to cut that overhead by proposing Lean code, reading checker feedback, and revising until the proof compiles. In other words, Leanstral treats formal verification tools as an interactive workspace, not a static target: it edits files, runs shell commands, and uses the Lean language server while working through real repositories. That design aligns with how developers already debug code, which is why this model feels like a practical assistant rather than a research demo.
Open-source access and cost: democratizing machine-checkable proofs
The most important decision here is not the parameter count; it is the licensing and access path. Leanstral 1.5 is a free Apache-2.0 open source AI code agent for Lean 4 verification, available both via Mistral’s Labs API and as downloadable weights on Hugging Face. Apache-2.0 licensing makes local inspection, fine-tuning, and downstream experimentation far more practical than a closed API-only release, which matters for researchers who need reproducibility and for companies wary of sending proprietary proofs to a black-box service. Mistral reports that Leanstral solves 587 of 672 PutnamBench problems and estimates a cost of about USD 4 (approx. RM18.40) per problem, compared with USD 300 (approx. RM1,380) or more for a rival system, Seed-Prover 1.5 high. For teams exploring formal methods on a budget, those economics, combined with USD 0 (approx. RM0) Labs pricing today, remove one of the biggest excuses not to experiment with machine-checkable proofs at scale.
From benchmark leaderboards to real bugs and long proofs
Benchmarks matter because they signal ceilings, and Leanstral 1.5 raises that ceiling for Lean 4 reasoning: it fully saturates miniF2F, solves 587 of 672 PutnamBench problems, reaches 87% on FATE-H and 34% on FATE-X, and lifts FLTEval pass@8 from 31.9 to 43.2. But the more interesting story is what happens off-benchmark. Despite being trained mainly on math, Leanstral has shown strong code verification abilities, discovering previously unknown bugs in open-source repositories, including 11 genuine bugs and 5 that had never been reported on GitHub before. One long-running proof attempt about AVL-tree insertion and deletion time-complexity guarantees consumed more than 2.7 million tokens across 22 context compactions, showing that the model can keep working on deep verification tasks rather than giving up after a few steps. For developers, this means the model is not only a solver for contest-style problems but a partner for messy, long-lived codebases.
A crowded prover race and what comes next for Lean developers
Leanstral 1.5 enters an active race rather than a vacuum: other Lean 4 theorem-proving efforts include Seed-Prover, DeepSeek-Prover, AlphaProof-based systems, the Goedel-Architect framework, Aleph Prover, and the OpenATP interface that already plans to update its integration to target Leanstral 1.5. The Labs listing makes clear that this particular endpoint is experimental, with retirement scheduled for September 30, 2026, so users should treat it as a fast-moving target rather than a fixed production dependency. Yet that timeline should push teams to try it now, not wait: repository evaluations have already surfaced 47 violated properties, 11 genuine bugs, and 5 previously unreported issues, and the next meaningful test will be reproduction on independent Lean codebases. If developers and researchers adopt Leanstral into their Lean 4 proof engineering workflows, the effect will be cumulative: more machine-checkable proofs in the open, better formal verification tools, and a community that treats correctness as a standard part of development, not a luxury add-on.
