Which Model for Reasoning: Small 4, Large 4, Medium 3.5

Lesson rewritten November 2, 2026. This lesson previously covered the Magistral models, all of which were retired on July 31, 2026. It now compares the models currently offering reasoning capabilities on the Mistral platform, based on their official specification sheets: Mistral Small 4 · Mistral Large 4 · Mistral Medium 3.5.

Three models, three profiles

At Mistral, reasoning is no longer the domain of a specialized lineup. Instead, three general-purpose models offer it, each with its own price point, context window, and maturity level.

Pricing and status data recorded November 2, 2026: verify these figures on the official specification sheets before calculating your budget.

Mistral Small 4: The starting point

Small 4 is described as a hybrid model that combines instruction following, reasoning, and coding in a single architecture. It is the default choice for this course: available in General Availability (GA) and low-cost, it switches from a direct response (none) to full reasoning (high) without changing models. Its specification sheet also lists function calling, structured outputs, agents and conversations, and batch processing. You can use it throughout an application, paying for reasoning only where it adds value.

Mistral Large 4: The most demanding cases

Large 4 is the most capable model in the lineup, featuring a 1-million-token context window and a three-tier reasoning scale. The documentation recommends high as a starting point, then suggests medium or low when token usage or latency becomes a priority. Keep two caveats in mind: it is in Public Preview, so it is subject to change, and it does not accept none. Reserve it for problems where Small 4 falls short, after benchmarking it on your own examples.

Mistral Medium 3.5: For existing deployments only

Medium 3.5 served as the official replacement for Magistral Medium and Devstral 2, and its documentation recommends reasoning_effort="high" for agentic workflows and coding tasks. However, it has been deprecated since October 31, 2026 and will be retired on April 30, 2027, with Large 4 designated as its successor. Do not start new projects on this model; if you are already using it, plan your migration now.

What about third-party models?

The Mistral platform also hosts models from other publishers. The reasoning page cites zai-glm-5-3, which accepts low, high, and max, but not none, and always returns a list of fragments including the reasoning trace. This illustrates the rule from Lesson 2: every model has its own set of valid reasoning_effort values.

The fate of the Magistral lineup

Magistral was Mistral’s first reasoning lineup. Its reasoning was always active, with no parameter to control it; it performed best with a default system prompt, adjustable via a prompt_mode parameter. All versions are now retired:

Since Medium 3.5 is itself deprecated, any migration from Magistral Medium should target Large 4 directly, or Small 4 if the quality suffices. Lesson 8 walks through this migration step by step.

Making a practical choice

Start with Small 4 set to high on a representative sample of your use cases. If the quality meets your needs, stick with it: it is by far the least expensive option. If not, test Large 4 with medium, then high. Each time, compare output quality, tokens consumed (usage.completion_tokens), and latency. Once you move to production, pin a dated identifier (mistral-small-2603 rather than mistral-small-latest): an alias can switch to a new version without your code changing, which may break your application.

Key takeaways

• Three models currently support reasoning: Small 4 (GA), Large 4 (Public Preview), and Medium 3.5 (deprecated)
• Small 4 is the starting point: $0.15 / $0.60 per million tokens, with high or none
• Large 4 offers low, medium, high and a 1-million-token context window, but no none
• All Magistral versions are retired; Medium 3.5 will be retired on April 30, 2027
• Choose based on measurement, then pin a dated identifier for production