Mistral Medium 3.5

Here's the thing about frontier models right now: most force you to choose between coding capability and pure reasoning, or charge you an arm and leg for both. Mistral Medium 3.5 is the first merged flagship to smash coding, reasoning, AND instruction-following into one set of weights, and it shows — 77.6% on SWE-Bench Verified puts it at Day Rank #10. The kicker is the configurable reasoning effort, meaning you can dial depth up when you need it and cut API costs down to $1.50/input, $7.50/output per million tokens when you don't. Open weights on HuggingFace means teams can self-host rather than eat proprietary lock-in, which for anyone building agentic pipelines is legitimately the move.