Language modelPreview

Mistral Large 4

Mistral AI

Released: October 6, 2026
Data date: October 7, 2026

Price: $1.36 · $4.18
Input · Output / 1M tokens
Input: Text, Image
Output: Text

Mistral Large 4 is Mistral’s largest model to date, trained from scratch on 3,800 Grace Blackwell GPUs in its own European data centers. This multimodal mixture-of-experts checkpoint accepts text and images. It has been available as a public API preview since October 6, 2026, with weights announced for the end of October.

Mistral particularly highlights the combination of visual grounding and agentic work. Its demos show Large 4 examining engineering drawings and large satellite images by zooming into regions and checking what it finds. For coding, Mistral reports 61.7% on DeepSWE v1.1. In a blind coding-quality evaluation conducted with Surge AI, the preview scored 3.74 out of 5, behind Claude Opus 5 and ahead of Kimi K3 and GLM 5.3. These results concern the preview; Mistral is continuing to train the model.

Specifications and access

Specification	Value and source
Parameters	1.05TSource
Notes	API preview. Weights are announced for the end of October 2026. The launch post lists 49 billion active parameters, while the model page lists 52 billion. Mistral advertises 1M tokens of context; OpenRouter documents 524,288 tokens.Source
Model type	Open weights announced; currently an API previewSource
Access	API (public preview)Source
Weights	Announced for the end of October 2026; not yet releasedSource
Active parameters	49 billion in the launch post; 52 billion on the model page. Difference unresolved.Source
Context by provider	1M tokens (Mistral); 524,288 tokens (OpenRouter)Source
OpenRouter context window	524,288 tokensSource
Maximum output on OpenRouter	262,144 tokensSource

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

Cybench (Show measurement, test conditions, and source)

Source value: 93
Score: 93
Metric: challenge success rate
Unit: %
Category: safety
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Security-competition tasks; the dataset contains 40 challenges, not a disclosed run sample size.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

DeepSWE (Show measurement, test conditions, and source)

Source value: 61.7
Score: 61.7
Metric: reported score
Unit: %
Benchmark version: 1.1
Category: coding
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Agentic software engineering; DeepSWE v1.1.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

SWE-Atlas-QnA (Show measurement, test conditions, and source)

Source value: 59.4
Score: 59.4
Metric: reported score
Unit: %
Category: coding
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Repository understanding and codebase question answering.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

Terminal-Bench (Show measurement, test conditions, and source)

Source value: 28.3
Score: 28.3
Metric: reported score
Unit: %
Benchmark version: 4.0
Category: coding
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Terminal-Bench 4; Mistral launch evaluation.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

Mistral Coding Agent Index (Show measurement, test conditions, and source)

Source value: 49.8
Score: 49.8
Metric: combined coding index
Unit: %
Category: coding
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Combined Coding Agent Index reported by Mistral, not a single benchmark or a Gradually overall score.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

Dense 200 (Show measurement, test conditions, and source)

Source value: 42
Score: 42
Metric: visual grounding score
Unit: %
Category: visual
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Visual grounding in dense natural scenes; Mistral's Dense 200 evaluation.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

B3 AI Security Benchmark (Show measurement, test conditions, and source)

Source value: 93.3
Score: 93.3
Metric: attack resistance
Unit: %
Category: safety
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Lakera public B3 benchmark; resistance to attacks.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

Cyber harmful-request refusal (Show measurement, test conditions, and source)

Source value: 95.3
Score: 95.3
Metric: average refusal rate
Unit: %
Category: safety
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Average refusal across cyber prompts from JailbreakBench, StrongREJECT and AgentHarm. This measures refusal of the tested harmful requests.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

SciCode-Verified (Show measurement, test conditions, and source)

Source value: 91.8
Score: 91.8
Metric: pass@1
Unit: %
Category: coding
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. SciCode-Verified pass@1; the chart states n=6 without defining whether that denotes repeated runs or another sample dimension.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

Mistral human STEM evaluation (Show measurement, test conditions, and source)

Source value: 69
Score: 69
Metric: weighted preference rate
Unit: %
Category: reasoning
Direction: Higher is better
Source type: vendor-reported
Evaluator: Mistral Large 4 launch
Status: active
Retrieved at: 2026-10-07
Methodology: Vendor-reported result. Mistral Large 4 public preview. Internal expert evaluation on mathematics and physics against GLM-5.3. The chart explicitly reports a weighted win rate of 69%; weighting is not specified.
Context: Mistral reports the public-preview model. Run date, effort, harness and evaluation sample size are not specified. Reproduced third-party results remain provider-reported here.

Source Mistral Large 4 launch

AutomationBench-AA (Show measurement, test conditions, and source)

Source value: 59.90136246808686
Score: 59.90136246808686
Metric: partial workflow score
Unit: %
Category: agentic
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Mistral Large 4 Preview, exact AA entry mistral-large-4. Partial-score business-workflow evaluation.
Context: Independent AA measurement for its exact Mistral Large 4 Preview entry. Raw embedded public model data is preserved; fractions are displayed as percentages. No run date or exact effort setting is provided on the inspected page. Mistral rounds this AA result to 59.9% in its launch post.

Source Artificial Analysis

Terminal-Bench (Show measurement, test conditions, and source)

Source value: 26.7676767676768
Score: 26.7676767676768
Metric: reported score
Unit: %
Benchmark version: 4.0
Category: coding
Direction: Higher is better
Source type: independent-evaluator
Evaluator: Artificial Analysis
Status: active
Retrieved at: 2026-10-07
Methodology: Independent evaluation. Mistral Large 4 Preview, exact AA entry mistral-large-4. AA's Terminal-Bench 4.0 run; distinct from the provider's 28.3% launch result.
Context: Independent AA measurement for its exact Mistral Large 4 Preview entry. Raw embedded public model data is preserved; fractions are displayed as percentages. No run date or exact effort setting is provided on the inspected page.

Source Artificial Analysis

Page 1 of 9

Sources and data date

Type	Evidence and data date
Price	Base API prices without caching or batch discounts. Higher context tiers and other rates are listed under Costs. Source: Pricing · Model metadata (retrieved October 7, 2026) · Mistral Large 4 model documentation · Editorial description reviewed October 7, 2026
Input	Documented modalities. Retrieved October 7, 2026. Source: Model documentation · Release announcement (retrieved October 7, 2026) · Mistral Large 4 announcement · Editorial description reviewed October 7, 2026
Output	Documented modalities. Retrieved October 7, 2026. Source: Model documentation · Release announcement (retrieved October 7, 2026) · Mistral Large 4 announcement · Editorial description reviewed October 7, 2026
Research status	Research date October 7, 2026. 31 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Unresolved	launch-human-pairwise-evaluation Original wording A further human-judged pairwise chart names peers and win rates but does not identify its task set or evaluation protocol; its values are not imported.
Unresolved	LiveBench aggregate scores and cost per successful task Original wording LiveBench also publishes overall and category aggregates for mistral-large-4-high, and its public page displays $0.147 per successful task. The official repository says aggregates are computed in the browser from task rows, but the checked dashboard and release assets do not establish that these views share one exact snapshot and aggregation path; the cost CSV separately records cost_per_successful_task=0.1385. These observations are therefore not merged with the raw subtask values, and no discrepancy or error is inferred.
Additional source	Context window data (retrieved October 7, 2026)
Additional source	Mistral October 2026 launch pricing (retrieved October 7, 2026)
Additional source	Artificial Analysis (retrieved October 7, 2026)
Additional source	Artificial Analysis (retrieved October 7, 2026)
Additional source	Artificial Analysis (retrieved October 7, 2026)
Additional source	Vals AI, Vals Index (retrieved October 7, 2026)
Additional source	Vals AI, Code Migration (retrieved October 7, 2026)
Additional source	Vals AI, Excel Modeling Benchmark (retrieved October 7, 2026)
Additional source	Vals AI, Finance Agent v2 leaderboard (retrieved October 7, 2026)
Additional source	Vals AI, Legal Research Bench (retrieved October 7, 2026)
Additional source	Vals AI, MedCode (retrieved October 7, 2026)
Additional source	Vals AI, MedScribe (retrieved October 7, 2026)
Additional source	Vals AI, MysteryMechanism (retrieved October 7, 2026)
Additional source	Vals AI, ProofBench (retrieved October 7, 2026)
Additional source	Vals AI, Tax Agent Bench (retrieved October 7, 2026)
Additional source	Vals AI, Vibe Code Bench 1-100 (retrieved October 7, 2026)
Additional source	Vals AI, Vibe Code Bench (retrieved October 7, 2026)
Additional source	Vals AI, BioMysteryBench (retrieved October 7, 2026)
Additional source	Vals AI, Harvey’s Legal Agent Benchmark leaderboard (retrieved October 7, 2026)
Additional source	Vals AI, IOI (retrieved October 7, 2026)
Additional source	Vals AI, Terminal-Bench (retrieved October 7, 2026)
Additional source	Arena WebDev leaderboard (retrieved October 7, 2026)
Additional source	LiveBench 2026-06-25, raw task-score CSV (retrieved October 7, 2026)
Additional source	LiveBench 2026-06-25, cost CSV (retrieved October 7, 2026)

Mistral Large 4

Specifications and access

Measurements without matching peer values

Our tests

Pelican on a bicycle

Published prices

Compare this model

Sources and data date