Mistral Medium 3

Mistral Medium 3, announced on 7 May 2025 under the slogan "Medium is the new large", is the model with which the French company put its pricing argument most bluntly. It is a closed-weights multimodal model with a context window of about 128,000 tokens, taking text and images in and producing text out, served under the API name mistral-medium-2505. The headline was the price: $0.40 per million input tokens and $2.00 per million output tokens, which Mistral described as roughly eight times cheaper than comparable models while claiming performance at or above 90% of Claude Sonnet 3.7 across its benchmark suite, ahead of Llama 4 Maverick and Cohere Command A. The second argument was deployment. Unlike most models in its class, Medium 3 was offered for hybrid, on-premises and in-VPC installation, self-hostable from four GPUs upwards, with continued pretraining and full fine-tuning available for enterprise customers in finance, energy and healthcare. The model was marked deprecated on 22 May 2026 and Mistral retires it on 31 August 2026, pointing users to Mistral Medium 3.5. That migration is expensive: the successor is priced at $1.50 input and $7.50 output per million tokens, so the recommended upgrade path costs 3.75 times as much per token in both directions. Mistral's own larger model, Mistral Large 3, is cheaper still at $0.50 and $1.50 — which makes Medium 3 the last point at which the company's mid-tier line was also its cheap line. Its direct catalogue sibling Mistral Medium 3.1 (v25.08) kept the same $0.40 / $2.00 pricing and retires on the same day.

▸News

Mistral retires Medium 3 in eight days — the replacement it names costs 3.75 times more per token

8/23/2026 · Mistral AI — cennik API i tabela wycofań w dokumentacji, komunikat premierowy Medium 3; zestawienie własne wujec.ai

▸Videos

No videos yet.

▸Reviews