Mistral Medium 3

Mistral AI announced Mistral Medium 3 on May 7, 2025 as a new efficiency-focused enterprise tier, positioning it at roughly 90% of Claude Sonnet 3.7 on benchmarks while pricing at $0.40 input and $2.00 output per million tokens. The launch post claims it surpasses leading open models such as Llama 4 Maverick and enterprise peers like Cohere Command A, and outperforms cost leaders such as DeepSeek v3 in both API and self-deployed setups. The blog frames the model as multimodal, coding- and STEM-strong, and deployable on any cloud including self-hosted environments starting at four GPUs, with features like hybrid or in-VPC deployment, custom post-training, and integration into enterprise tools. Pricing, deployment footprint, and benchmark claims all originate from Mistral's own announcement rather than independent verification, so they should be read as creator-stated positioning for the Medium 3 variant.

Mistral Medium 3 is a proprietary multimodal large language model released by Mistral AI on May 7, 2025, positioned as a mid-tier enterprise offering. It accepts both text and image inputs and produces text outputs, operating over a 128,000-token context window that supports long documents and extended conversations. The model can be self-hosted on as few as four GPUs, making it relatively accessible for enterprise-grade deployment. Mistral launched Medium 3 with pricing of $0.40 per million input tokens and $2.00 per million output tokens, marketing it as delivering roughly 90% of frontier model performance. At launch the API was available through Mistral La Plateforme and Amazon SageMaker, with subsequent availability announced for IBM watsonx, NVIDIA NIM, Azure AI Foundry, and Google Cloud Vertex AI. The model was followed by a Medium 3.1 refresh in August 2025 and served as the base for Magistral Medium, Mistral's first reasoning model, in June 2025.

Mistral Medium 3 was released by MistralAI on May 7, 2025, as a versatile language model designed for efficient text and image processing. It supports text, image, and file inputs while producing text outputs, with a context window of 131,072 tokens suitable for long-context enterprise workflows. The model is optimized for coding, STEM reasoning, and enterprise adaptation, and supports hybrid, on-prem, and in-VPC deployments across major cloud environments. The model exposes standard API parameters including frequency_penalty, max_tokens, presence_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, and top_p, with defaults of temperature 0.3 and top_p 0.95. Its model ID is mistralai/mistral-medium-3, reflecting its route through Mistral's catalog and third-party gateways. The directory frames Medium 3 as a balanced performance option with reduced operational costs for scalable professional applications.

Epoch AI's independent registry entry confirms Mistral Medium 3 as a closed-weights model from Mistral AI, released May 7, 2025 in France, with an ECI score of 134 that ranked 146th of 257 tracked models overall and 29th of 136 at launch. Its per-benchmark scorecard reports DTBench at 62%, MATH Level 5 at 82%, OTIS Mock AIME at 32%, GPQA Diamond at 60%, and Lech Mazur Writing at 7.7, offering the most granular independent capability data in the candidate set. The page mirrors creator pricing of $0.40 input and $2.00 output per million tokens and frames the model as delivering state-of-the-art performance at 8X lower cost with simplified enterprise deployment. Training parameter count, compute, hardware, and dataset size remain unknown, and headline figures ultimately trace back to Mistral's own launch claims rather than Epoch's independent measurement.

Mistral Medium 3 (Premier v25.05) is documented on Mistral's official model page as a frontier-class multimodal release dated May 7, 2025, with a 128k context window and support for text plus image inputs. The page enumerates production capabilities including structured outputs, function calling, document QnA, prefix caching, chat completions, agents, batching, and predicted outputs across /v1/chat/completions, /v1/conversations, /v1/agents, and /v1/batch endpoints. The same documentation now marks Mistral Medium 3 as deprecated with a deprecation date of May 22, 2026, directing integrators to the sibling Mistral Medium 3.5 for new work. The model's stated lifecycle status is the most consequential update from the page for anyone currently routing traffic to this variant, and the explicit replacement naming disambiguates it from related Medium family builds.

OpenOCR publishes reproducible benchmark evidence for Mistral: Mistral Medium 3 on the corpusocr-shared-11-normalized/v2 corpus using the ocr-shared-11-scorer/v2 scorer, measured 2026-08-22 via the openrouter provider offering. Across 11 documents the model achieves an average score of 97.4% with an average latency of Per-document results include 100.0% on a synthetic grocery receipt (4.3s) and 100.0% on a cursive handwritten letter (1.2s) and handwritten notes (1.6s), with a café receipt scoring 86.7% at 2.9s. The page exposes the exact input image, ground truth, and returned transcript for each test case, with a corpus manifest SH

OpenRouter's listing for Mistral Medium 3 (mistralai/mistral-medium-3) describes it as a high-performance enterprise-grade language model positioned at frontier-level capabilities at reduced cost, balancing reasoning and multimodal performance and excelling in coding, STEM reasoning, and enterprise adaptation. The page The same OpenRouter page enumerates three Mistral-hosted provider variants — Mistral (ZDR) at $0.40/$2.00 per 1M with a $0.04 cache read rate and 0.37s P50 latency at 41 tps and 99.56% uptime, standard Mistral at the same listed price with 0.34s P50 latency and 47 tps throughput at 99.53% uptime, and Mistral (EU) price