Mistral Medium 3.5 Review

Mistral Medium 3.5 arrived on 30 April 2026 with a consolidation story that is becoming the theme of the year. Rather than keep separate reasoning and coding specialists, Mistral merged its Magistral and Devstral 2 lines into a single 128 billion parameter dense model with a dial it calls reasoning effort. The result is a compact, frontier leaning generalist that scores 77.6 percent on SWE-Bench Verified and still fits on four GPUs.

Value is the through line. The API runs about 1.50 dollars per million input tokens and 7.50 dollars per million output, the context window stretches to 262,144 tokens, and vision and function calling come as standard. Because the weights ship under a modified permissive licence, teams that want to own the stack can pull it in house and run it on modest hardware, which is a rare combination at this level of coding skill.

What keeps it out of the very top tier is exactly what makes it efficient. A dense 128B model trails the giant sparse flagships on the hardest reasoning, and Mistral is a smaller shop than the hyperscalers behind the leaders. But for coding, agentic work and high volume automation where cost and control matter more than the last few benchmark points, Medium 3.5 is a very smart buy.