Skip to content
Best AI Ever

Mistral AI · AI model

Mistral Large 4

Mixture-of-Experts model with 1.05T total and 52B active parameters, in Public Preview since Oct 6, 2026. Text and image input, 1M-token context window and adjustable reasoning. Open weights are announced but not yet released.

Last verified Preview 9 sources

Key specs

Provider
Mistral AI
Family
Mistral Large
API identifier
mistral-large-4-0
Availability
Preview
Release date
2026-10-06
License
Proprietary
Context window
1,000,000 tokens
API access
Yes
Consumer app
Unknown

Capabilities

Inputs

  • Text input Yes
  • Image input Yes Technical documentation lists image input maximum size as 8
  • Audio input No Technical documentation marks audio input as not supported (N/A)
  • Video input No Technical documentation marks video input as not supported (N/A)
  • File input Yes PDF documents via Document QnA (OCR + model) in Chat Completions/Conversations; Document QnA is listed on the model page; uploaded files max 50 MB and 1,000 pages

Outputs

  • Text output Yes
  • Image output No Only text output listed; image generation exists separately as a built-in agent tool
  • Audio output No Only text output listed

Features

  • Reasoning Optional Adjustable via reasoning_effort parameter ('none' or 'high')
  • Tool calling Yes Function Calling listed for /v1/chat/completions and /v1/conversations
  • Web search Yes Built-in web_search / web_search_premium tools via the Conversations and Agents APIs (not Chat Completions); model page lists Built-In Tools

"Unknown" means we have not verified this capability yet; it does not mean the model lacks it.

Pricing

Estimate your monthly cost →
DimensionTierPriceSinceVerifiedSource
Input tokens base $1.36 per 1M tokens 2026-10-12 2026-10-25 API pricing | Mistral Docs Standard mode list price (Priority mode +75% and regional inference +10% not included). The 50% launch discount ('Launch pricing: 50% off for 2 weeks.', changelog Oct 6, 2026) has ended: the docs pricing page shows no sale price as of 2026-10-25
Cached input tokens base $0.14 per 1M tokens 2026-10-12 2026-10-25 API pricing | Mistral Docs Cached input price listed on the docs pricing page (10% of the input price); launch discount ended
Output tokens base $4.18 per 1M tokens 2026-10-12 2026-10-25 API pricing | Mistral Docs Standard mode list price; launch discount ended
Batch input tokens base $0.68 per 1M tokens 2026-10-12 2026-10-25 API pricing | Mistral Docs Batch API: 50% of standard price (docs pricing page Batch mode, factor 0.5: 'Batching ... at a 50% discount'; also stated on mistral.ai/pricing); computed from the standard list price
Batch output tokens base $2.09 per 1M tokens 2026-10-12 2026-10-25 API pricing | Mistral Docs Batch API: 50% of standard price (docs pricing page Batch mode, factor 0.5); computed from the standard list price

Prices are shown in the currency the provider publishes; we never convert currencies.

Benchmarks

BenchmarkSettingScoreReported byMeasuredVerifiedSource
DeepSWE v1.1 % resolved (pass@1), higher is better v1.1, 'Mistral Large 4 Preview'; effort not stated; Artificial Analysis harness ('AA-DeepSWE 1.1', run privately by AA), not mini-swe-agent 61.7 % Self-reported — 2026-10-25 Introducing Mistral Large 4 | Mistral
GDPval-AA v2.1 Elo, higher is better v2.1, 'Mistral Large 4 Preview'; effort not stated; Stirrup harness (shell + web browsing) 1,422 Elo Third party 2026-10-25 2026-10-25 GDPval-AA v2.1 Leaderboard (Artificial Analysis)
Terminal-Bench 4.0 % tasks resolved, higher is better Terminal-Bench 4.0, 'Mistral Large 4 Preview'; effort not stated; Artificial Analysis harness (run privately by AA before the harness' public launch) 28.3 % Self-reported — 2026-10-25 Introducing Mistral Large 4 | Mistral
Text Arena Arena score, higher is better Overall, style control (default view); no effort suffix 1,434 Third party 2026-10-23 2026-10-25 Text Arena leaderboard (Arena)
DeepSWE v1.1 caveats:

The official leaderboard uses mini-swe-agent for all models and shows whole-number percentages with margins of about ±1 to ±6 points. Providers often report scores from their own scaffolds or effort settings; DeepSeek reports 65.5% to 74.2% for the same model depending on the scaffold. DeepSWE v1 scores are not comparable with v1.1. With only 113 tasks, small differences are within noise.

GDPval-AA v2.1 caveats:

Elo is a relative rating: values change as models are added and the scale is re-anchored, so always store the date a rating was read. For example, Anthropic's launch page gives 1846 for Claude Opus 5.5, while the leaderboard on 2026-10-12 shows 1866 (Max, Default Fallback). Earlier versions (Google cites "GDPVal-AA v2") are not comparable. Judging is done by an LLM, not by human experts. Anthropic models are listed with "Default Fallback", meaning refused steps are routed to a fallback model. AA-Briefcase and GDP.pdf are separate evaluations.

Terminal-Bench 4.0 caveats:

Scores depend heavily on the agent harness (for example Claude Code, Codex, Grok Build or mini-SWE-agent) and on the reasoning effort, so the same model appears several times on the leaderboard. Leaderboard rows are often submitted by the organisation that builds the agent. Providers' own runs can differ from the leaderboard: Anthropic reports 66.4% for Claude Opus 5.5 at max effort, while the leaderboard lists 64.8% ± 3.1%. Versions 2.0, 2.1 and 3.0 and Artificial Analysis' "Terminal-Bench Hard" use different task sets and are not comparable with 4.0. Margins are about ±3 points.

Text Arena caveats:

Scores measure voter preference, which can be influenced by answer style and length. The leaderboard offers a style-control adjustment, so record which view was used. Scores shift with every update, so a value is only meaningful together with its leaderboard date. Entries marked "Preliminary" or with few votes have wide confidence intervals. The same model can appear several times with different reasoning-effort suffixes (for example -high or -max). Category views (Coding, Math, Creative Writing and others) are different rankings from Overall.

One row per result, without rank numbers: each result was measured under its own setting (effort level, harness, tools, test version), as stated by the source, and scores compare only within the same setting. Self-reported results come from the provider and have not been reproduced independently.

Rankings

Derived from our published ranking snapshots; each ranking explains its criteria and weights.

Compare Mistral Large 4

Guides

About Mistral Large 4

Mistral Large 4 is a general-purpose multimodal Mixture-of-Experts model with 1.05T total parameters, 52B active parameters and a 1.6B vision encoder. It launched as a Public Preview on October 6, 2026 and is available through the Mistral API as mistral-large-4-0 (alias mistral-large-4). It accepts text and image input, has a 1M-token context window and supports adjustable reasoning via the reasoning_effort parameter. Mistral has announced that the weights will be released by the end of October 2026.

Strengths and weaknesses

Strengths

  • 1M-token context window (other current Mistral models in this catalog: 256k)
  • Text and image input with adjustable reasoning (reasoning_effort)
  • Supports structured outputs, function calling, Document QnA, batch processing and built-in tools (Agents/Conversations APIs)

Weaknesses

  • Public Preview: may receive silent updates and has no guaranteed path to General Availability (model lifecycle policy)
  • Weights not yet published (model page lists weights as 'Coming soon')
  • No audio or video input

Sources

  1. Mistral Large 4 - Mistral AI | Mistral Docs
  2. Mistral Large 4 - Technical Documentation For Downstream Providers
  3. Document QnA | Mistral Docs
  4. Reasoning | Mistral Docs
  5. Websearch | Mistral Docs
  6. API pricing | Mistral Docs
  7. Introducing Mistral Large 4 | Mistral
  8. GDPval-AA v2.1 Leaderboard (Artificial Analysis)
  9. Text Arena leaderboard (Arena)

Last verified Page updated 2026-10-27