Magistral models

Strongest: Magistral Medium 1.2

Magistral Medium 1.2 (alias magistral-medium-2509, shipped 18 September 2025) is Mistral's flagship reasoning model: a Premier/closed-weight, multimodal model that produces visible chain-of-thought before answering, with a 40,960-token output ceiling for long traces. It posts AIME24 91.82% / AIME25 83.48% and GPQA Diamond 76.26%, and its reasoning traces stay in the user's language — a genuine multilingual-reasoning differentiator. Priced at $2.00/$5.00. The buyer's sentence: the right pick for auditable, multilingual hard-reasoning routing — though Medium 3.5's reasoning knob has narrowed its exclusive niche. - Provider: Mistral AI (Paris, France) - Release: 2025-09-18, status GA - Context: 131,072 tokens; max output 40,960 (long thinking traces) - Modalities: text + image in, text out (native multimodal) - Knowledge cutoff: ~June 2025 - Headline price: $2.00 input / $5.00 output per 1M tokens - Architecture: dense (parameter count undisclosed — Premier/closed)