Mistral launches low‑cost code‑proofing agent, pushes DIY AI to challenge OpenAI and

Leanstral is now available as an agent mode in Mistral Vibe and via a free API endpoint, with its weights released under Apache 2.0, according to The Register. The company says the model, Leanstral‑120B‑A6B, outperforms larger open‑source rivals such as GLM5‑744B‑A40B, Kimi‑K2.5‑1T‑32B and Qwen3.5‑397B‑A17B on its internal FLTEval benchmark for engineering proofs.

The Register reports Leanstral’s pass@2 score of 26.3 beats Anthropic’s Sonnet by 2.6 points while costing only $36 to run, versus Sonnet’s $549. At pass@16, Leanstral reaches 31.9, an 8‑point margin over Sonnet, the firm claims. The cost advantage is framed as a “high‑value alternative to the Claude suite,” per Mistral’s own data.

Mistral’s broader strategy targets enterprise customers who want to control their AI pipelines. TechCrunch notes the launch of Mistral Forge, a platform that lets firms train custom models from scratch on proprietary data, positioning the company against OpenAI and Anthropic. CEO Arthur Mensch says the move is helping Mistral approach $1 billion in annual recurring revenue this year.

VentureBeat adds that Mistral is expanding its AI stack with products like OCR 3 for document digitization and a new coding assistant aimed at GitHub Copilot’s market share. Together with Leanstral, these tools signal a push to offer end‑to‑end, low‑cost AI solutions for software development and data‑intensive enterprises.