Mistral Large 4

Oct 2026
Mistral Large 4
Compare in
on
$1.13per task?
$1.36 / $4.18per 1M in / outcache $0.14
67 tok/sfirst token in 1.7 s?
1Mcontext

Rankingin thinking mode?

Full ranking →
1Sonnet 5.51,817.32Opus 5.51,8033Muse Wonder Full1,729.34Fable 5.11,672.35Opus 51,659.5
···
29Qwen 27B1,394.930Mistral Large 41,39231Sonnet 51,358.4

Strong at

A
Coding#23 of 132Agents & work tasks#37 of 175Reasoning#62 of 334Long context#33 of 273
B
Knowledge#84 of 318Vision#39 of 124

Alternatives?

ModelScoreCostPer taskWhat changes
CheaperClaude Sonnet 5.5highwins 80%-22%$0.88No forced tool use+5BetterMiMo v2.6 FlashthinkingOpenwins 64%-94%$0.062No PDF input+4OpenQwen3.8 Flashthinkingwins 75%-67%$0.37Context: 1M → 262K+6CheaperClaude Haiku 5.5maxwins 74%-81%$0.21Remove sampling parameters+5CheaperDeepSeek V4.1 Flash 1017maxOpenwins 74%-77%$0.26No PDF input+5

History

4 versions
5001,0001,500Jan 26Apr 26Jul 26Oct 26Release dateAA-Briefcase 1.134
5001,0001,500Jan 26Apr 26Jul 26Oct 26Release dateAA-Briefcase 1.134

Benchmarks

in Agents & Tool Use
Terminal-Bench 4.026.8#23
GDPval-AA (work tasks) 2.146.1#37
AA-LCR (long context) 1.181.3#33
AutomationBench-AA59.9#19
AA-Briefcase 1.1 · Elo1,392#30
Harvey LAB-AA 1.13.1#13
DeepSWE 1.1lab62
AutomationBenchlab59.9
Finance Agent v2lab54.7
Harvey Legal Agent Benchmarklab15.8
CyberGym-E2E AAlab82
Cybenchlab93
Finch FinWorkBenchlab67.4
B3 Attack Resistancelab93.3
Harmful cyber request refusal meanlab95.3
AA Briefcase combined Elolab1,393
KORA aggregatelab1.7

Watch out 5

Infinite whitespace stream›

In JSON Object mode, include JSON in the system or user prompt to avoid a possible infinite whitespace stream. Source ↗

Content is a list of chunks›

With reasoning_effort high, message.content is a chunk list rather than a plain string; parse thinking and text chunks. Source ↗

Only function calling · Mistral regional endpoints›

Regional endpoints support function calling only and omit Agents, Batch and Files; check regional model availability before routing requests. Source ↗

10% upcharge · Mistral regional endpoints›

Regional inference charges 1.1× standard list prices for input, output, cached reads and cache writes. Source ↗

Reinforcement learning run still in flight›

The public preview reinforcement learning run is still in flight, and Mistral expects further rapid changes as training scales. Source ↗

FAQ

How much does Mistral Large 4 cost?›

$1.36 per 1M input tokens and $4.18 per 1M output tokens; cached input $0.14 per 1M; cheapest via mistralai at $0.68 / $2.09. A typical benchmark task costs about $1.13 at default effort.

Is Mistral Large 4 free to use?›

There is no free API; the cheapest paid price is $0.68 per 1M input tokens.

How good is Mistral Large 4, and what is it best at?›

#62 of 334 in Chat & Knowledge, #23 of 132 in Coding, #19 of 133 in Agents & Tool Use, #90 of 195 in Writing & Language, #39 of 124 in Vision & Documents. Strongest at: Coding, Agents & work tasks, Reasoning, Long context.

Is Mistral Large 4 better than Mistral Large 3?›

On AA-Briefcase 1.1 · Elo, Mistral Large 4 scores 1,392 vs 228.3 — higher. It is 3,544% more expensive per task.

How can I use Mistral Large 4?›

Through the API of Mistral.

What is the context window of Mistral Large 4?›

1M tokens, with up to 262K output tokens.

When was Mistral Large 4 released? Is there a newer version?›

Released 2026-10-06. It is the current version in its line.

Is Mistral Large 4 open source?›

No, Mistral Large 4 is a closed model available through an API. The closest open-weights option here is Qwen3.8 Flash.

Can Mistral Large 4 read images?›

It accepts text and images, PDF files as input.

What is Mistral Large 4?›

Mistral Large 4 is a closed language model from Mistral, released Oct 2026, the successor to Mistral Large 3.

What are the best alternatives to Mistral Large 4?›

Claude Sonnet 5.5 (cheaper), MiMo v2.6 Flash (better), Qwen3.8 Flash (open weights).