Mistral AI: models, pricing and agents
Mistral AI is a French lab that sells access to its models through an API and publishes some of them as open weights. Its lineup covers small models (Ministral), general-purpose models (Small, Medium, Large), code (Devstral, Codestral) and voice, and every Mistral model in our dataset is listed with tool-calling support. To build an agent, the choice depends less on the brand than on the exact model, its price and its reliability on your tasks.
Who Mistral AI is and what it sells
Mistral AI is a Paris-based company that has published language models since 2023. It offers three entry points: an API for developers, the consumer assistant Le Chat, and downloadable models when the license allows it. Prices and the model list are published on the official pricing page, which prevails if it differs from our data.
To place the offer in the wider market, our model directory lists more than 350 models from all vendors, with the same reading grid: context, price, tool calling, open weights.
The lineup in one table
The figures below come from our 10/24/2026 snapshot (source: OpenRouter). Prices are in dollars per million tokens, input and output. They change: each model page is refreshed with every snapshot.
| Model | Input / output price | Context | Open weights | Reasoning |
|---|---|---|---|---|
| Mistral Large 4 | $1.36 / $4.18 | about 1 million tokens | no | yes |
| Mistral Medium 3.5 | $1.50 / $7.50 | 262,144 tokens | yes | yes |
| Mistral Small (March 2026) | $0.15 / $0.60 | 262,144 tokens | yes | yes |
| Devstral (Dec. 2025) | $0.40 / $2.00 | 262,144 tokens | yes | no |
| Ministral 8B | $0.15 / $0.15 | 262,144 tokens | yes | no |
| Codestral (Aug. 2025) | $0.30 / $0.90 | 256,000 tokens | no | no |
Three points stand out. Small models cost 7 to 28 times less than Mistral Large 4 on output. A newer model is not always priced higher than an older one: Mistral Large 4 is cheaper on output than Mistral Medium 3.5 in this snapshot. Finally, the “open weights” column changes from one row to the next: “Mistral” does not mean “open source”. To understand what that label really guarantees, read our guide on open vs closed models for your agents and the definition of open weights.
What an agent costs on Mistral
An agent re-reads its context at every step (instructions, history, tool results), so the bill depends mostly on input tokens, repeated at each turn. Here is a worked example to adapt to your numbers: a 20-step task with 15,000 input tokens per step and 500 output tokens per step.
- Total input: 20 × 15,000 = 300,000 tokens; total output: 20 × 500 = 10,000 tokens.
- With Mistral Small ($0.15 / $0.60): $0.045 + $0.006, about $0.05 per task.
- With Mistral Large 4 ($1.36 / $4.18): $0.408 + $0.042, about $0.45 per task.
The 9x gap carries over to a full month of work. Prompt caching lowers the repeated input when the provider offers it (the snapshot lists a cache price for several Mistral models). To estimate your own case with your volumes, use the LLM price calculator, and read our analysis of the real cost of an AI agent for the costs beyond the model.
Mistral inside an agent: tool calling and coding agents
Tool calling is the feature that lets a model request an action (read a file, query an API). Every Mistral model in our dataset shows it as supported, but listed support does not guarantee reliability: a model may fill in tool parameters badly or loop. The LLM Benchmark crosses scores and prices, and our guide to reading the leaderboards explains their limits.
On the product side, Mistral offers an agents API and a command-line coding agent, Mistral Vibe, described in the official documentation. You can also plug a Mistral model into an open source agent that accepts several providers, such as Aider, Cline or Goose, through an API key or a local model. To compare families of coding agents, see our article on coding agents.
Running it at home
Open-weight models in the lineup can run on your own machine. The small ones (Ministral 3B or 8B, Mistral Small) need a consumer graphics card or a recent Mac once quantized; the larger ones need servers. The VRAM calculator estimates the memory required depending on quantization, and the Ollama tutorial shows how to launch a model locally. The benefit: your data stays on your machine, which matters for GDPR (the EU data protection regulation); the trade-off is hardware to buy and maintain.
Who it suits, and who it does not
- Good fit: teams that want a European provider, small and cheap models for well-defined tasks, or the option to go local with open weights.
- To verify: the model’s behavior on your real agent tasks (tools, long contexts), the license of the version you download, and where data is processed if you have contractual obligations.
- Less suited: when a tool requires a specific model (some agents are tuned for a single provider) or when you need a plug-in ecosystem that only exists at another vendor.
What to know before you choose
- Compare by model, not by vendor: the price gap between two Mistral models often exceeds 10x.
- “Open weights” is a property of each model, not of Mistral in general.
- Test 10 to 20 real tasks from your agent before migrating.
- Re-read the official pricing page on the day you decide: our figures date from 10/24/2026.
- For a broader view of the choice, see our guide to choosing an LLM.
Frequently asked questions
Is Mistral AI open source?
Only in part. Mistral AI publishes some open-weight models (for example Mistral Small, Ministral or Devstral in our dataset) and keeps others available only through its API (Mistral Large). Before any commercial use, read the license of the exact version you download.
Which Mistral model should I pick for an AI agent?
For an agent, first look for tool-calling support, then the price-to-reliability ratio. A small or mid-size model is often enough for well-defined tasks; a large model is justified for long reasoning or complex code. Test on your own tasks before deciding.
Is Mistral AI cheaper than ChatGPT?
On several lines, the per-million-token prices we recorded for Mistral are lower than those of the leading US models, but the gap depends on the model compared and on cache discounts. Compare with the price calculator on your real volume, not on a list price.
Can I run Mistral locally?
Yes for open-weight models, with a tool such as Ollama and enough video memory. Closed models are only available through the API or the provider's own offers.