Mistral Medium 3.5 has credible upper-mid-tier developer capability: Mistral’s card identifies a frontier-class multimodal model for coding and agents, while Artificial Analysis reports a 30 Intelligence Index and a 256k context window. Its 77.6% SWE-bench Verified result is vendor-reported, not an independently reproduced leaderboard result. This places it below GPT-5.5 and Claude Opus 4.8 on demonstrated frontier capability, but above GLM-4.6’s technical baseline and close to the lower end of Kimi K2.5’s range. Coding evidence is mixed but usable. The supplied Terminal-Bench aggregator reports 50.6% on Terminal-Bench 2.1 (rank 54/142), supporting competent terminal work but not the leading agentic tier; DeepSWE and LiveCodeBench provide no listed result. Accordingly it trails DeepSeek-V3.2 and Kimi K2.5 in coding/agentic confidence despite the SWE-bench claim. Official multimodal positioning supports a materially higher I/O score than text-oriented DeepSeek-V3.2 and GLM-4.6, though modality-level quality details are absent. At $1.50/M input and $7.50/M output, it is materially cheaper than premium frontier anchors but not a budget leader. Artificial Analysis’ 76.3 output tok/s and 2.16s TTFT support an above-median speed score. Mistral documentation and API materials support solid developer experience. Public preference is moderate: Text Arena rank 87, 1427±7, across 11,019 votes shows meaningful usage but substantially weaker preference than leading published anchors.
