Choosing the Right Model

A Model for Every Need

With over a dozen models in the catalog, choosing one can feel intimidating—and we often see this roadblock stop beginners in their tracks. They spend more time hesitating between models than actually using them. This lesson gives you a simple decision framework to select the right Mistral model for your situation, based on three factors: performance, cost, and availability. The guiding principle is simple: start from your need, never from the model.

The Decision Tree

The fundamental question is: what do you want to do? Let’s walk through the most common cases.

If you need to converse or write—Q&A, drafting, summarizing, translating—start with Mistral Small 4. It’s fast, versatile, and its 256k token context window handles very long documents. Only upgrade to Mistral Medium 3.5 (the frontier model) or Mistral Large 3 if results lack precision or depth—both come with open weights.

If you need to code, the answer depends on the task’s ambition. For code completion in an IDE, Codestral is the natural choice. For more complex software engineering tasks—multi-file refactoring, feature implementation—choose Devstral 2. For demanding code agents, Mistral Medium 3.5 is also built for code and agentic workflows.

If you need to reason, remember the previous lesson: reasoning is now built into generalists. Mistral Small 4’s reasoning mode covers moderately complex problems—math, logic, analysis. For the most demanding cases, step up to Mistral Medium 3.5. The Magistral family, once dedicated to this use case, is deprecated.

If you process documents, distinguish two situations. To extract text from images or scanned PDFs, OCR 4 is the solution. To analyze and synthesize already textual documents, Mistral Small 4, Medium 3.5, or Large 3 work perfectly.

Finally, if you work with voice: Voxtral TTS to convert text to speech, Voxtral Mini Transcribe Realtime to transcribe speech to text in real time.

Comparison Table of Main Models

Speed and cost indications are relative within the range; for exact rates per million tokens, check the official Mistral pricing page.

Open vs Premier: How It Affects Your Choice

You noticed the “Tier” column in the table. This distinction has very concrete consequences for your project. Open models are accessible via the Mistral API, but also downloadable: you can host them on your own servers, which eliminates API costs for self-hosting and gives you total freedom of use and modification. This applies to Mistral Medium 3.5, Mistral Small 4, Mistral Large 3, Devstral 2, and Ministral 3. For an organization whose sensitive data must not pass through an external service—an HR department processing employee files, for example—this possibility changes the nature of the choice.

Premier models, conversely, are accessible only via the Mistral API: weights are not downloadable, billing is pay-per-use (tokens consumed), and you benefit from Mistral’s technical support. Codestral, OCR 4, and Moderation 2 fall into this tier. If data sovereignty is your priority, prioritize Open models that you can host internally.

The Cost Factor

API usage cost relies on two variables that multiply: the chosen model—larger, more powerful models cost more per token—and the number of tokens consumed per request, input and output combined. The general rule we recommend is simple: always start with the cheapest model that meets your quality requirements, and move up only if results are insufficient. Doing the opposite—starting at the top “to play it safe”—means paying frontier rates for tasks a compact model would handle just as well. For high-volume applications, like a consumer-facing chatbot or batch processing, Ministral models offer the best performance-to-cost ratio.

Methodology for Choosing

In summary, here is the systematic approach to apply to every new project:

1. Identify your need: conversation, code, documents, voice, reasoning?
2. Evaluate complexity: a simple question or multi-step analysis?
3. Consider volume: a few requests per day or thousands per hour?
4. Check constraints: internal hosting required? Limited budget?
5. Test with Small 4: it’s the best starting point in 80% of cases
6. Adjust: move up to Medium 3.5 or Large 3 if quality is insufficient, down to Ministral if cost is too high

This test-adjust loop beats any theoretical table: your real use case decides.

Pricing as of August 5, 2026 — prices change regularly: before any budget calculation, check the current rate card on the official Mistral pricing.

Key Takeaways

• Start with Mistral Small 4 in most cases—it’s fast, versatile, and economical
• Mistral Medium 3.5 is the fallback when quality comes first (frontier, agents, demanding code)
• Use specialized models when your task justifies it (code, voice, documents)
• Open models offer self-hosting flexibility
• Test and iterate: the best model is the one that delivers satisfactory results at the lowest cost