Mistral Medium 3 fits engineering teams that want near-flagship coding help without flagship pricing, especially ones able to self-host on 4 GPUs and above for VPC control. It has already been superseded twice, so treat it as a budget option for existing deployments, not a new long-term integration.
Mistral Medium 3 is Mistral AI's mid-tier multimodal large language model, accepting text and image input and scoring 0.953 on DocVQA, a document visual question-answering benchmark. It handles OCR-heavy documents and codebases directly, positioned as a lower-cost alternative to larger proprietary flagship models for teams needing multimodal reasoning without frontier-scale compute.
Where it sits
- $0.8/M$ per 1M tokensBlended price (3:1)Lower is better#20 / 60peer median $1.86/Mvendor price, checked by HokAI
- 37 tok/stokens/sOutput speedHigher is better#33 / 34peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 26peer median 78.3%per source, see benchmark scores
- 57.1%% correctGPQA DiamondHigher is better#39 / 42peer median 86.9%per source, see benchmark scores
Cheaper than 68% of the 60 GA models with a published price, in the bottom third on GPQA Diamond (rank 39 of 42), and one of 63 whose vendor states it does not train on customer data. Ranked against GA models; this record is not GA.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Mistral AI · Family: Mistral Medium
More about Mistral AI on HokAI
Context window: 128,000 tokens
Input modalities: text, image · Output: text
About Mistral Medium 3
Mistral Medium 3 (API name mistral-medium-2505) is a frontier-class multimodal model released by Mistral AI on May 7, 2025. It sits in the middle of Mistral's lineup, positioned below the flagship Large tier but designed to deliver most of the capability of larger proprietary models, including Claude Sonnet 3.7, at roughly an 8x lower price per token. Mistral has not disclosed the exact parameter count or whether the model uses a dense or mixture-of-experts architecture, and pitched the launch under the tagline "medium is the new large." On benchmarks, Mistral Medium 3 scored 92.1% on HumanEval 0-shot, outperforming Llama 4 Maverick's 85.4%. On Math500 Instruct 0-shot it reached 91.0%, and on MMLU-Pro 5-shot with chain-of-thought it scored 77.2%, both ahead of GPT-4o's reported figures on the same tests. On ArenaHard 0-shot it reached 97.1%. The model's weakest reported area is graduate-level reasoning: GPQA Diamond 5-shot CoT came in at 57.1%, well behind frontier reasoning-tuned models released later in 2025 and 2026. Mistral did not publish SWE-bench, AIME, or ARC-AGI scores at launch, a transparency gap that drew criticism from independent evaluators. The model's context budget is large enough to hold full codebases and lengthy multi-file documents in a single request; a separate maximum output token limit was not published, so output tokens draw from that same budget. On RULER, a long-context retrieval benchmark, it scored 0.902, ahead of GPT-4o's reported 0.889 at the same context length. Independent evaluation by Artificial Analysis recorded an output speed of 36.8 tokens per second and a time-to-first-token of 1.53 seconds against the hosted La Plateforme endpoint. Mistral Medium 3 accepts text and image input and produces text output; on DocVQA it scored 0.953, and on MMMU it scored 0.661. The model supports function calling and tool use, structured JSON outputs, fill-in-the-middle completions, document OCR and Q&A, and Mistral's Agents and Conversations APIs. It is a non-reasoning model, with no extended-thinking or configurable reasoning-effort mode, unlike Medium 3.5's later "reasoning effort" parameter. Mistral Medium 3 is available via Mistral's La Plateforme API and Amazon SageMaker at launch, with Amazon Bedrock, Azure AI Foundry, Google Cloud Vertex AI, IBM watsonx, and NVIDIA NIM added as deployment targets since. Enterprises that need on-premises or VPC control can self-host on 4 GPUs and above, with support for continuous pretraining and fine-tuning. Mistral AI is headquartered in Paris and operates under EU data protection law by default; its published policy is that API inputs are not used to train future models unless a customer opts in, and Mistral has not published a dedicated system card with HarmBench or jailbreak-resistance figures for Medium 3 specifically. Mistral shipped two successors within a year: Medium 3.1 in August 2025, and Medium 3.5, a larger dense model with native reasoning-effort controls, in April 2026. Medium 3 carries a published deprecation timeline around May 2026 on at least one cloud listing, so new integrations should target Medium 3.5 while existing mistral-medium-2505 deployments plan a migration.
Pricing
$0.40 per 1M input tokens and $2.00 per 1M output tokens on Mistral's La Plateforme API, Amazon Bedrock, and Azure AI Foundry, roughly an 8x discount versus comparable proprietary flagship pricing at launch. Artificial Analysis lists a blended price near $0.56 per 1M tokens using a typical 7:2:1 input:output:cache weighting. Self-hosting is available for teams that want to avoid per-token billing, at infrastructure cost instead of a subscription fee.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.012 | $0.0020 | $0.014 |
| Support reply | $0.0008 | $0.0006 | $0.0014 |
| One coding agent run | $0.080 | $0.040 | $0.120 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Long Context Window: Ships with a 128K-token context window, per Mistral's own model card, enough headroom to send an entire small codebase in one request.
- Native Vision and Multimodal Understanding: Accepts image input alongside text and scored 0.661 on MMMU, a multimodal reasoning benchmark, without a separate OCR pipeline.
- Long-Context Retrieval Accuracy: Scored 0.902 on RULER, a long-context retrieval benchmark, ahead of several proprietary flagship models at the same window size.
- Function Calling and Agents API: Supports tool use, structured JSON outputs, fill-in-the-middle completions, and Mistral's Agents and Conversations APIs for multi-step workflows.
- Self-Hostable Deployment: Runs in a customer's own VPC or on-premises environment for teams that can't send data to a third-party API, with support for continuous pretraining and fine-tuning.
Pros
- Delivers flagship-tier coding accuracy at a fraction of typical proprietary API pricing, making it a strong pick for budget-conscious coding assistants.
- Strong at document understanding and long-context retrieval, useful for OCR-heavy and multi-document workloads.
- Deployable via managed API or self-hosted in a customer's own VPC, giving enterprises a path to keep data in-house.
Cons
- GPQA Diamond score of just 57.1% trails frontier reasoning models on hard science questions.
- Output speed of 36.8 tokens per second is well below the roughly 94.5 tok/s median for comparable models, a poor fit for latency-sensitive chat.
- Superseded twice within a year by Medium 3.1 and Medium 3.5, with a deprecation timeline around May 2026, so new integrations should target the newer model instead.
Benchmarks
- MATH: 91% vendor-reported · 07 May 2025 — Competition maths problems, % solved.
- Mmmu: 66.1 vendor-reported · 07 May 2025
- MMLU-Pro: 77.2% vendor-reported · 07 May 2025 — A harder version of the 57-subject knowledge exam, % correct.
- HumanEval: 92.1% vendor-reported · 07 May 2025 — Small programs that must pass hidden tests, % passing.
- Arena Hard: 97.1 vendor-reported · 07 May 2025
- Ruler 128k: 90.2 vendor-reported · 07 May 2025
- GPQA Diamond: 57.1% vendor-reported · 07 May 2025 — PhD-level science questions that are hard to search for, % correct.
- AA blended price: $0.56/M cited: Artificial Analysis · 14 Jun 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
- Output speed: 37 tok/s cited: Artificial Analysis · 14 Jun 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
What are Mistral Medium 3's pricing plans in 2026?
Mistral Medium 3 costs $0.40 per 1 million input tokens and $2.00 per 1 million output tokens, available through La Plateforme, Bedrock, or Azure AI Foundry. A document OCR and QA pipeline processing 500K input and 50K output tokens a day would run about $0.30, while a coding-agent loop at 2M input and 400K output tokens a day would run about $1.60. Neither figure includes any cached-input discount, since Mistral hasn't published one for this model.
Can you use Mistral Medium 3 without paying?
Mistral Medium 3 has no free tier. It's priced entirely per token from the first request, with no trial credits or free monthly quota published by Mistral. The only way to try it without a subscription is a pay-as-you-go setup on La Plateforme, where a small test workload costs well under a dollar.
What are Mistral Medium 3's closest competitors?
Mistral Medium 3's closest hokai-tracked alternatives are GPT-4o, OpenAI's 2024 multimodal model, and Mistral's own Large 3, a bigger 675B open-weight MoE model with a 256K context window. For dedicated document extraction rather than general chat and coding, Mistral OCR 4 is a narrower, purpose-built alternative. GPT-4o adds real-time voice Medium 3 lacks, Large 3 trades proprietary pricing for open weights and more context, and OCR 4 trades general capability for deeper document-extraction features.
How does Mistral Medium 3 compare to GPT-4o in 2026?
Mistral Medium 3 outperforms GPT-4o's reported 90.2% HumanEval score, while GPT-4o adds real-time voice input and output that Medium 3 doesn't support. Both models are already deprecated by their makers in favor of newer releases, so neither is the pick for a brand-new long-term integration. For self-hosting or lower per-token pricing, Medium 3 is the more flexible of the two, since GPT-4o is API-only through OpenAI and Azure.
How long does it take to get going with Mistral Medium 3?
Getting started means creating a Mistral API key on La Plateforme, or using it through Amazon Bedrock or Azure AI Foundry if you're already on one of those clouds. The API accepts text and image input, supports structured JSON output and function calling out of the box, and uses the model ID mistral-medium-2505. Teams needing on-premises or VPC deployment instead of the hosted API can self-host, though that requires a commercial agreement with Mistral rather than a simple sign-up.
Top Alternatives
- GPT-4o: Pick Mistral Medium 3 for higher HumanEval accuracy and self-hosting; pick GPT-4o if you need real-time voice, since Medium 3 has no audio support.
- Mistral Large 3: Pick Mistral Large 3 for open weights and a 256K context window; pick Mistral Medium 3 for a smaller, cheaper proprietary API.
- Mistral OCR 4: Pick Mistral OCR 4 for dedicated document extraction with bounding boxes; pick Mistral Medium 3 for a general-purpose multimodal model that also handles chat and coding.