Back to the index
Mistral AI · Mistral
No. 23

Magistral Medium 1.2

Mistral's transparent-chain reasoning flagship, now multimodal.

ReasoningClosed weightsText, Image
Field notes
Context window
131Ktokens
Max output
41Ktokens
Input price
$2/ 1M tokens
Output price
$5/ 1M tokens
Modalities
Text, Image
Knowledge cutoff
Jun 2025
Released
Sep 2025
Approx. speed
Overview

What this model is

Magistral Medium 1.2 is Mistral's proprietary reasoning model, producing explicit multilingual chains of thought and strong performance on math and structured problem-solving, now with a vision encoder for reasoning over images. It is competitive on reasoning benchmarks and notably good at showing its work in the user's language, but its agentic coding reliability remains below the closed frontier. Best used for math, logic, and analysis rather than autonomous engineering.

Strengths

  • Dedicated, transparent reasoning traces
  • Strong competition-math performance (AIME)
  • Multimodal reasoning over text and images
  • Function calling and JSON mode support

Trade-offs

  • Slower and pricier per task (long thinking output)
  • Closed weights (Magistral Small is the open variant)
  • Agentic coding below the frontier
  • Narrower general knowledge than flagships

Best for

  • Complex math and scientific reasoning
  • Multistep logic and analysis
  • Visual reasoning tasks
  • Auditable chain-of-thought workflows

Not ideal for

  • Latency-sensitive or high-volume chat
  • Long autonomous coding agents
Capabilities

Capability profile

Normalized 0–100 scores, comparable across the whole catalog.

Reasoning80Coding66Math84Writing72Knowledge76Speed58Agentic62Vision60Multilingual80Long Context72
Reasoning
80
Coding
66
Math
84
Writing
72
Knowledge
76
Speed
58
Agentic
62
Vision
60
Multilingual
80
Long Context
72
Benchmarks

How it scores

Public benchmark results, with independent third-party results where available. Bars normalize percentages to 100 and Elo ratings to a 1500 ceiling.

Independent index
Artificial Analysis Intelligence Index
Composite of ~9–10 independent evals · Artificial Analysis
18/ 100
BenchmarkResult
Artificial Analysis Intelligence Indexindependent
GeneralArtificial Analysis
18%
Aider Polyglotindependent
CodingAider
47.1%
Alternatives

Comparable models

Models in a roughly similar class — worth weighing against this one.