Experiment and Scale Plans: Quotas and Pricing
On this page
Two plans for two stages of a project
Mistral offers two API access plans, each designed for a different phase of development. Understanding the difference prevents a classic nasty surprise: an application that works perfectly during a demo but collapses the moment thirty users connect simultaneously.
The Experiment plan is free. It provides access to all Mistral models and is intended for prototyping, learning, and evaluation. You can test Mistral Small, Mistral Large 3, and Devstral 2, debug your integration, measure response quality on your real-world use case, and build a working prototype. In return, it enforces strict rate limits—a maximum number of requests per second, per minute, and per month, varying by model. Once you exceed them, the API returns 429 Too Many Requests. This isn’t an outage; it’s the plan reminding you of its purpose.
The Scale plan is the professional tier, billed based on usage. It offers significantly higher quotas, priority support, and access to advanced features. Activating it is surprisingly discreet: in the Mistral Console, go to Settings > Billing, add a payment method, and the switch happens instantly. Your existing API keys continue to work exactly as before, now with the new quotas—no code changes, no redeployment required.
| Feature | Experiment (Free) | Scale (Production) |
|---|---|---|
| Cost | Free | Pay-as-you-go (per million tokens) |
| Available models | All models | All models |
| Rate limits | Restricted (1–2 req/s depending on model) | High (adjustable upon request) |
| Monthly quota | Limited (capped tokens/month) | Unlimited (based on budget) |
| Support | Documentation + community | Priority support |
| SLA | None (best-effort) | Guaranteed SLA (availability, latency) |
| Recommended usage | Prototyping, learning, POC | Production, commercial applications |
| Quota increase | Not available | Upon request via support |
How billing is calculated
Scale billing is based on the number of tokens consumed. A token corresponds to approximately four characters in English, and slightly fewer in French, where accents and longer words break into more pieces: for equal content, a French text therefore costs slightly more than an English one. You are billed separately for input tokens—your prompt, the context you attach, the conversation history—and for output tokens, meaning the generated response. Output tokens generally cost more because they require more compute resources.
Four optimization levers follow from this mechanism, and they are more effective than any commercial negotiation. The first is model choice: Mistral Small is much cheaper than Mistral Large 3, and a classification or extraction task has no need for the most powerful model in the catalog. The second is the max_tokens parameter, which caps response length—leave it open, and the model will write three paragraphs where one sentence would have sufficed. The third is reducing the context sent: in a multi-turn conversation, only relevant history deserves to be retransmitted, since it is billed on every call. The fourth is application-side caching: if the same ten questions return daily in your FAQ, store the answers on the application side rather than regenerating them.
The Console gives you the tools to monitor all of this without waiting for the invoice. You will find token volume by model and by day, cumulative costs for the current period, configurable email alerts when you approach a threshold you defined, and spending limits that cut off usage before costs spiral. Configure these last two the same day you activate Scale, not the following month.
When to switch
Stay on Experiment as long as you are exploring: the plan is made for that, and it costs you nothing. Switching to Scale becomes necessary when real users access your application, when you hit requests-per-second limits, when you need an SLA and reachable technical support, or when your application becomes critical and can no longer tolerate interruptions. In practice, these four signals often arrive together, within the week following a successful first demo in front of a client.
Pricing as of August 5, 2026 — prices change regularly: before any budget calculation, check the current rate card on the official Mistral pricing page.
Key takeaways
- Experiment = free, ideal for learning and prototyping, limited quotas
- Scale = pay-as-you-go, designed for production, high and adjustable quotas
- Billing is based on tokens (input + output), with different rates per model
- Configure spending alerts as soon as you activate the Scale plan
- Moving from Experiment to Scale is transparent: your existing API keys remain valid