What happened
Mistral AI released Leanstral, an open-source code agent for Lean 4, under an Apache 2.0 licence. Designed for formal proof engineering, Leanstral (6B active parameters) aims to reduce human review bottlenecks in high-stakes software. Benchmarked on FLTEval, Leanstral-120B-A6B achieved a 26.3 score at pass@2 for $36, outperforming Sonnet's 23.7 score at $549. It also surpassed larger open-source models in efficiency, scaling to 29.3 at pass@4 for $72. The model is available in Mistral Vibe, via a free API, and as downloadable weights.
Why it matters
Leanstral significantly lowers the cost barrier for formal code verification, enabling platform engineers to integrate proof-based assurance into development workflows. Its $36 cost for pass@2 performance, surpassing Sonnet's $549, shifts unit economics for high-integrity software. This mechanism reduces reliance on expensive human expertise for critical codebases, potentially accelerating deployment timelines. This follows Anthropic's recent focus on AI code security tools, intensifying competition in verified code generation. The Apache 2.0 licence and downloadable weights provide procurement teams with deployment flexibility and control.
Subscribe for Weekly Updates
Stay ahead with our weekly AI and tech briefings, delivered every Tuesday.
