Leanstral 1.5: Mistral’s AI Found Five Real Bugs

Mistral released Leanstral 1.5, a formal verification agent built on Lean 4, which found five previously unreported bugs in open-source repositories during its first public test. The tool mathematically proves code correctness for all possible inputs, catching a critical overflow bug in a Rust crate that could silently corrupt data in release builds. Mistral claims the system reduces the cost of formal verification from hundreds of dollars to roughly $4 per proof, making it accessible for everyday software development.

Every AI coding tool released in the last two years has chased the same goal: generate more code, faster. Mistral just shipped something different. Leanstral 1.5 https://mistral.ai/news/leanstral-1-5/ , released July 2, is a formal verification agent built on Lean 4. In its first public test against 57 open-source repositories, it found five bugs that human maintainers had never reported. The Bug That Makes the Case The most striking find was in datrs/varinteger https://crates.io/crates/varinteger , a Rust crate for variable-length integer encoding. In the zigzag decoding function, the sign operation performs a value + 1 calculation — fine for most inputs, until you feed it the maximum u64 value. In debug mode, the overflow panics. You know something is wrong. In release builds — the ones that ship — the overflow wraps silently, corrupting output with no diagnostic warning. That is the kind of bug that survives code review, passes test suites, and only surfaces when production data gets quietly mangled. Leanstral caught it not by running tests, but by formally proving what the function can and cannot do. What Formal Verification Actually Means Formal verification is not debugging, fuzzing, or property-based testing. It is mathematical proof that a property holds for all possible inputs — not a large sample, not edge cases you thought to test. All of them. Lean 4 https://lean-lang.org/theorem proving in lean4/ is both a programming language and a theorem prover. Write your code, state your property as a type, and Lean’s kernel verifies the proof. If it compiles, the property is guaranteed. The verdict is binary: proven or not. There is no “usually correct.” Tests can show a bug exists. Formal proofs show it cannot. The Cost Curve Changed Everything Formal verification used to mean hiring mathematicians and spending weeks on a single module. That economics made it realistic only for aerospace, medical devices, and cryptographic protocols where failure costs are catastrophic. Leanstral 1.5 solves Putnam Mathematical Competition problems at roughly $4 per proof. Competing systems charge $300 or more. It saturated miniF2F 100%, the first model to do so and solved 587 of 672 PutnamBench problems at 4 million reasoning tokens. The model scales with thinking time: 44 problems at 50k tokens, 587 at 4M. The API is free during beta. The weights are on Hugging Face https://huggingface.co/mistralai/Leanstral-1.5-119B-A6B under Apache 2.0. Commercial use is permitted. How to Start Today The fastest path is Mistral Vibe: uv tool install mistral-vibe vibe --setup vibe --agent lean For tighter Lean integration, add the Lean LSP MCP server to ~/.vibe/config.toml : mcp servers name = "lean-lsp" transport = "stdio" command = "uvx" args = "lean-lsp-mcp" The model is also available at the labs-leanstral-1-5 API endpoint — OpenAI-compatible clients work — and self-hostable on vLLM 0.24.0+. Architecture: 119B parameter mixture-of-experts with 6B active per token. Efficient enough to run at reasonable cost even if you bring your own compute. The Fair Critique Hacker News was quick to note that the varinteger overflow is exactly what property-based testing catches. Fair. proptest or cargo-fuzz explore boundary values like u64::MAX by default. But the critique misses what formal verification actually offers. The value is not catching this specific bug — it is proving that an entire class of inputs is handled correctly, permanently. Property-based testing narrows likelihood. Formal proofs eliminate possibility. Leanstral makes the stronger tool accessible at developer economics for the first time. The Shift Underway 2025 was the year AI learned to write code. 2026 is shaping up to be the year AI learns to guarantee it. AWS already uses Lean 4 to formally verify Cedar https://aws.amazon.com/about-aws/whats-new/2023/05/cedar-open-source/ , its authorization policy language. Cryptographic libraries are moving the same direction. Safety-critical teams are paying attention. When formal verification runs from your terminal for dollars rather than from a consultancy for months, the question stops being can we afford to verify this? and starts being which parts of our codebase deserve it first? For Rust crate authors, teams shipping AI-generated code, and anyone building software where silent failures are unacceptable — that question is now worth asking.