Mistral Medium 2604
Mistral AI
At a glance
- Price per 1M tokens (input)
- $1.50
- Price per 1M tokens (output)
- $7.50
- Context
- 262,144 tokens
- Free access
- yes (see below)
Refreshed daily; data verified August 15, 2026. Published August 12, 2026.
Mistral Medium 2604 is designed for teams that need a single model to handle multimodal inputs, structured workflows, and complex problem-solving without juggling multiple specialized systems. It suits tasks like document analysis with embedded images, automated data extraction followed by API actions, and multi-step reasoning for audit trails or technical support. The vendor's approach emphasizes reliability through explicit step-by-step reasoning and native tool calling, which lets the model invoke external functions directly rather than merely suggesting code. This makes it a practical choice for production environments where traceability and deterministic behavior matter more than raw creative output.
Specifications & pricing
| Input (per 1M tokens) | $1.50 |
|---|---|
| Output (per 1M tokens) | $7.50 |
| Context window | 262,144 tokens |
| Max output | 262,144 tokens |
| Capabilities | images, tool calling, reasoning |
LiteLLM community dataset (MIT), verified August 15, 2026. Official Mistral AI pricing.
What Mistral Medium 2604 would cost on your workload — run it through the cost calculator →
Where to try Mistral Medium 2604 for free
- Mistral AI offers a free chat — Le Chat (free plan). A vendor's free chat may run a different model from the same family — the exact model is not guaranteed.
Frequently asked questions
What is this model best used for in a business context?+
It excels at tasks that combine visual and textual information, such as reviewing scanned invoices or engineering diagrams, then taking action via integrated software. It also shines in scenarios requiring transparent logic, like compliance checks or troubleshooting guides, where you need the model to show its work step by step.
How do I get started with Mistral Medium 2604?+
You can access it through the Mistral platform or via an API endpoint. For tool calling, you define a set of functions in your code, pass their schemas to the model, and it will return structured requests to invoke those functions. Start with a small pilot project, such as a document summarization pipeline that also extracts key fields, to test integration before scaling.
How does this model differ from other models in the Mistral line?+
Unlike smaller or text-only siblings, this version adds native image understanding and robust function calling out of the box. It also places a stronger emphasis on explicit reasoning traces, meaning it will output intermediate steps for complex problems rather than jumping to a final answer. That makes it more suitable for regulated industries or debugging-heavy development, though it may be heavier to run than lighter options.
What are the main limitations I should plan around?+
The model requires careful prompt design to get the most reliable tool-calling behavior, especially when multiple functions are available. It can occasionally misinterpret ambiguous images, so human verification is recommended for high-stakes visual analysis. Also, its step-by-step reasoning increases response length and latency, which may not suit real-time chat applications with strict speed requirements.
Compare with others
Org chart: how to move your company onto AI
A practical map: which company roles and processes AI agents can take over, where to start, and in what order to roll it out.