Mistral Medium 3

Overview

Mistral Medium 3 is a proprietary language model from Mistral AI, announced on 7 May 2025 under the slogan "Medium is the new large"^[1]. The slogan states the product's positioning: a mid-sized model that, in Mistral's account, does the work of larger systems at a fraction of the cost. Unlike open-weight releases, it is offered as a commercial API and as a model that customers deploy on their own hardware.

Price and performance claims

Mistral set the API price at $0.40 per million input tokens and $2 per million output tokens^[1]. A token is the unit in which models read and write text, roughly a fragment of a word; input tokens are the prompt and context sent to the model, output tokens the text it produces. The two are billed separately because generating text costs more compute than reading it.

Mistral said the model performs at or above 90% of Claude Sonnet 3.7 across benchmarks at about 8x lower cost, and that it surpasses Llama 4 Maverick and Cohere Command A^[1]. Benchmarks are standardised test sets used to compare models on particular tasks; a score summarises accuracy on those tasks rather than guaranteeing how a model behaves in a specific workflow. According to Mistral, the model is strongest in coding and STEM tasks^[1].

Deployment

Mistral said Medium 3 can be run hybrid, on-premises or inside a VPC, on as few as four GPUs, and that enterprises can continue pretraining it or fine-tune it on their own data^[1]. A VPC — virtual private cloud — is an isolated network inside a cloud provider, so data stays within an organisation's own boundary. Mistral also said the model can be deployed on any cloud, including self-hosted environments of four GPUs and above^[1]. At launch it was available on La Plateforme, Mistral's own developer platform, and on Amazon SageMaker^[1].

The four-GPU floor is the part that matters commercially: a model that fits on a handful of accelerators can be self-hosted by a single team, which is out of reach for frontier-scale systems.

Versions

The first Mistral Medium, tagged mistral-medium-2312, was released on 11 December 2023^[3][4]. Mistral Medium 3.1 (mistral-medium-2508) followed on 12 August 2025^[2]. The build mistral-medium-2505 was deprecated on 22 May 2026^[3], and Medium 3 and 3.1 were retired on 31 August 2026, replaced by Mistral Medium 3.5^[3]. The date-like suffix in each model name is a snapshot identifier: the API points at a fixed build so that behaviour does not change silently under a customer.

Where it sits

Positioning an inexpensive mid-tier model as the default is a claim about where demand is: many applications do not need the largest model available, and paying for one is waste. Whether a 90% figure holds for a given task is something users have to test themselves. What the release made concrete was the packaging — a low per-token price, a small hardware footprint, permission to fine-tune, and availability on a third-party cloud platform at launch^[1].