Mistral launches 128B Medium 3.5 model, pricing draws criticism

Mistral AI released Medium 3.5 on April 29, a dense 128‑billion-parameter model distributed with open weights. The company set usage prices at $1.50 per million input tokens and $7.50 per million output tokens.

The release included the model weights, a cloud-based remote coding tool called Mistral Vibe CLI and a new Work Mode in Le Chat that runs multi-step autonomous tasks such as email triage, research synthesis and cross-tool coordination.

Mistral presented Medium 3.5 as a single set of weights that replaces three prior variants-Medium 3.1, Magistral and Devstral 2-and offers configurable reasoning effort per request. The company said the consolidation is intended to simplify engineering and deployment for customers.

On internal benchmarks, Mistral reported a 77.6% score on SWE-Bench Verified, which tests whether a model can generate working patches to fix real GitHub issues, and a 91.4% score on τ³-Telecom, a test of agentic tool use in specialized environments. Third-party leaderboard placements and independent evaluations were not available at publication.

Competing open models include Alibaba’s Qwen 3.6, Zhipu AI’s GLM and Xiaomi’s MiMo-V2. Qwen 3.6 is a 27‑billion-parameter model that scored 72.4% on SWE-Bench Verified and is distributed under an Apache 2.0 license that allows free download and self-hosting. Public leaderboards currently show the top positions held by those models.

Some developers and researchers questioned Medium 3.5’s per-token economics relative to smaller open models. University of Washington professor Pedro Domingos wrote on social media: “Regular AI companies brag about how much better their model is on benchmarks. Only Mistral brags about how much worse its one is.” Youssof Altoukhi wrote that political and regulatory factors had helped the company survive. Michal Langmajer wrote that he welcomed a non-U.S., non-Chinese lab building large models but called for stronger technical results and more competitive pricing.

Other observers framed open weights as a factor for buyers that need to run models on their own infrastructure. Industry sources said Mistral has secured enterprise deployments in Europe and that one bank signed a multi-year deal to self-host Mistral models. Mistral is headquartered in the EU.

Mistral described Vibe CLI as able to run parallel coding sessions and push pull requests to GitHub without a user at the terminal. The company said Work Mode automates multi-step workflows beyond single-turn chat.

Medium 3.5’s pricing and performance metrics will be available for wider assessment once independent benchmark results and additional customer deployments are published.