Magistral Deprecated: Migrating Reasoning

Update November 2, 2026. All Magistral versions were withdrawn as of July 31, 2026, like Devstral 2. Magistral Medium’s designated replacement, Mistral Medium 3.5, was itself deprecated as of October 31, 2026 in favor of Mistral Large 4; Mistral Small 4 remains the replacement for Magistral Small. The table below reflects status as of November 2; the AI Reasoning with Mistral course has been completely revamped around reasoning_effort.

The official models page now lists Magistral variants among deprecated models, with withdrawal dates ranging from November 30, 2025 to July 31, 2026 depending on the version. For anyone who built on Mistral’s reasoning range, this is the most structural change of the year: this is not a pricing change or an upgrade, but the planned end of an entire model family. This lesson explains why, then walks you through the migration end-to-end.

Why Mistral is withdrawing its reasoning range

This depreciation is not a reversal: it follows the logic described throughout this course. Reasoning stops being a separate model to become a capability integrated into generalists. Mistral Small 4 is explicitly described as unifying “instruct, reasoning, and code” in a single model; Medium 3.5 handles demanding agentic use cases. The direct consequence: maintaining a range dedicated to reasoning no longer has technical justification, since the capability now lives within the models you already use for everything else. For you, this simplifies architecture—one less model to choose, bill, and monitor—but it requires a migration that is better done before the withdrawal dates than after.

The mapping table

Here is the starting point for each affected model—and not just Magistral, since several other ranges are switching at the same time:

As always with this kind of table, a caveat applies: it gives the starting point, your evaluations make the decision. An official mapping never guarantees that the replacement model will behave identically on your prompts and your data.

Conducting the migration, step by step

The process consists of four steps, to be executed in order.

Start by inventorying the models called in your code and configurations: search for the strings magistral, mistral-large, devstral, and ocr. Pay particular attention to -latest aliases—they can switch silently to a new version without your code changing, which means your production may already depend on a model you have never tested.

Next, consult the withdrawal dates on the official models page, which lists them by version. For Magistral and Devstral, they have all passed: migration is no longer something to plan, it is urgent. This information sets your actual timeline.

Then comes the core of the work: testing integrated reasoning. Replay your Magistral cases on Small 4 first, then on Large 4 if the results are not sufficient (Medium 3.5, deprecated, would only be a step). Compare quality, but also compare cost: the generalist may prove more economical than the old specialist, and it happens that a forced migration results in a reduced bill. This is the kind of good surprise you only discover by measuring.

Finally, take advantage of the work to pin dated versions in production, instead of -latest aliases. You are touching every model call anyway: this is the ideal time to eliminate a source of silent switches, and avoid living through this situation again by discovering a behavior change after the fact.

Our affected courses

The “AI Reasoning with Mistral” course was revamped on November 2, 2026: it now teaches reasoning_effort on Small 4 and Large 4 and dedicates a lesson to migration from Magistral. The “Guide to Mistral Models” and “Codestral & Devstral” courses should be read with the mapping table above at hand.

Official source: docs.mistral.ai — Models Overview

Key takeaways

• Magistral was withdrawn as of July 31, 2026: any remaining calls must migrate
• Reasoning goes through reasoning_effort on Small 4 and Large 4 (Medium 3.5 being deprecated)
• Check the withdrawal dates for YOUR versions on the official page
• Pin dated versions in production during the migration

Test your knowledge

The recomposed 2026 range: verify that the map is up to date in your head.

The 2026 rule: a recomposed range, reasoning integrated into generalists—the migration lesson above gives the practical path.