Mistral Large 4: Features, API Pricing and Limitations

What is Mistral Large 4?

Mistral Large 4 is a multimodal AI model in public preview. Mistral's changelog dates the preview to October 6, 2026 and lists a 1M-token context window. It is worth evaluating for document-heavy or tool-assisted work, but a preview announcement does not establish reliability in your application. Official changelog.

This guide is for developers, agencies and founders deciding whether to run a controlled trial, not whether to replace their current stack immediately. We reviewed public documentation on October 7; we did not call the API, benchmark the model or test a deployment. Links are ordinary official-source links, not affiliate links.

Features and the parameter-count discrepancy

The model card lists 1.05T total parameters, 52B active parameters and a vision encoder. However, Mistral's launch announcement says 1T total and 49B active. These official sources disagree; we cannot reconcile the counting method. Use the model card for the current specification, but keep that discrepancy visible rather than declaring either number independently proven. Model card, launch announcement.

Total parameters describe the full model; active parameters describe the portion engaged during computation in this mixture-of-experts design. Neither number alone tells you the bill, response speed or whether your task succeeds. A founder buying an API integration should care more about accepted outputs and failure recovery than a large parameter headline.

The card lists structured outputs and function calling. Those are useful interfaces to test, not guarantees of valid application behavior. Model capabilities.

What a long context window does not prove

A large advertised window does not demonstrate that every fact in a long document will be retrieved correctly. Our recommendation is to test a short source set first, then deliberately place conflicting dates or missing pages in longer cases. Ask for source locations and check them against the originals.

For invoices, contracts or research, a polished summary is not an acceptance test. Decide in advance which fields, quotations and omissions matter. Do not upload client-sensitive documents until your account, processing terms and access controls have been checked.

API pricing: introductory versus listed rates

The card displayed these USD rates per million tokens on October 7:

Official pricing on the model card. The changelog describes 50% off for two weeks. Verify the actual billing terms before committing a client quote; we have not tested caching eligibility or an invoice. Discount announcement.

A worked estimate, not a measured bill

Assume one hypothetical request uses 10,000 uncached input tokens and 2,000 output tokens. Using the displayed rates:

• introductory estimate: (10,000 ÷ 1,000,000 × $0.68) + (2,000 ÷ 1,000,000 × $2.09) = $0.01098;
• listed-rate estimate: the same calculation with $1.36 and $4.18 gives $0.02196;
• 1,000 identical requests would therefore be $10.98 or $21.96 in this simplified example.

This arithmetic excludes retries, additional tool turns, retrieval, hosting, taxes, foreign-exchange charges and other billable usage. It assumes those input/output rates apply to the entire request. It is not a promise about your account's final charges.

For US and Indian teams, keep the provider's USD estimate separate from any local-currency customer quote. Do not present a rupee conversion as fixed without specifying the exchange rate and date.

Are the weights available to download?

The launch announcement plans weights for the end of October; the changelog says they are coming soon. A planned open-weight release is not a verified downloadable release on October 7. We have not inspected a weight package or licence. Launch plan.

If your requirement is immediate private hosting, check the actual files, licence and deployment requirements before choosing the model. Do not assume a hosted preview demonstrates that it will run economically on your own hardware.

What tool calling means for an agent

Mistral's function-calling guide separates model-generated arguments from developer-executed functions. The application executes the tool; the model's request is not itself an authorization. Function-calling documentation.

Our recommended boundary: start with read-only lookups. Validate tool names, input types, account scope and destinations in the application. Keep client sends, purchases, deployments and record deletion behind approval. Treat text retrieved from a webpage or file as data, not new instructions granting permissions. See our AI agent security checklist.

A four-stage evaluation before switching models

This is a proposed pilot, not a demonstration we ran.

1. Define a small test set. Choose twenty synthetic or approved tasks: source comparison, structured extraction, a code issue and a refused tool action. Write expected answers and a stop condition. Stop if the team cannot identify what counts as correct.
2. Run matched requests. Compare the same sources and permissions with your existing model. Record model ID, date, input/output usage, elapsed time, retries and result. Stop when a supposed comparison changes the task or tool access.
3. Review exact outputs. Check citations, schemas, calculations and code tests. Include a missing-source case and a malicious instruction embedded in retrieved content. Stop on unsupported critical facts or an unauthorized action; do not hide these failures inside an average score.
4. Price accepted work. Divide total model spend, including failed attempts, by outputs that pass your written criteria. Add human correction time and infrastructure separately. Keep the current model as fallback until failure handling is demonstrated.

For illustration, $0.60 of model spend across a trial with 18 accepted outputs gives about $0.0333 per accepted output. That is invented arithmetic for planning, not an observed Mistral result.

Who should evaluate it—and who should wait?

An agency can test document extraction without granting sending rights. A developer can test structured tool requests in a sandbox. A startup can compare quality and cost on its own workload. These are proposed uses, not customer case studies or revenue forecasts.

Wait if you require independently validated performance, verified downloadable weights today or production approval that your team has not completed. We are not assigning a GPT-versus-Claude-versus-Mistral winner; no independent benchmark was reproduced for this guide. For a broader starting point, see AI tools for freelancers.

FAQ

Is Mistral Large 4 free?

The hosted API has displayed token prices. Do not confuse an open-weight release plan with free hosted inference, free infrastructure or an unrestricted licence.

Is it already downloadable?

Not established by this review. The checked sources describe a future weight release. Verify the actual release before claiming self-hosting is available.

Is the active parameter count 49B or 52B?

The announcement says 49B; the model card says 52B. We preserve the conflict and have not independently verified its explanation.

Can it automatically update my CRM?

A compatible tool interface may let your application request an update. Permissions, validation, execution, logging and approval still need implementation. A model feature is not a ready-made CRM integration.

The practical decision

Run a small, scored trial if the model matches a real problem. Choose it only when accepted outputs, permission controls and total operating cost meet your criteria. Recheck preview terms and prices before production use or quoting a long-term project.

Sources

• Mistral Large 4 model card
• Mistral changelog
• Mistral launch announcement
• Mistral function calling