AI Agent StoreMISTRAL · OPEN WEIGHTS AVAILABLE · PREVIEW
Mistral Large 4
See how this model performed in tests and what it could cost for your tasks.
Thinking setting tested: high.
Data reviewed · LiveBench question set: 25 June 2026 · View every source ↗ · Download the data ↓
Add to a custom comparison ↗This result uses the exact LiveBench high configuration on the June 25 question set. The current routed endpoint declares 1,048,576 context and 262,144 maximum output tokens; these are serving limits. No exact Arena match or same-task none/high sweep is verified.
Results by type of task
Explain these tests ↗| Reasoning | 83.9 / 100 |
|---|---|
| Coding | 77.2 / 100 |
| Agentic coding | 57.2 / 100 |
| Mathematics | 93.6 / 100 |
| Data analysis | 76.5 / 100 |
| Language | 49.6 / 100 |
| Instruction following | 64.8 / 100 |
What the evidence says
Mistral Large 4 scores 71.8 on LiveBench, averaging its seven task categories equally. Its highest measured category is mathematics (93.6); its lowest is language (49.6). Categories differ in difficulty, so these are test results, not universal strengths or weaknesses.
Our General agent profile produces an Agent Fit score of 72.0 by emphasizing reasoning, instruction following, agentic coding, coding and data analysis. See the exact weights or adjust the profile in the interactive comparison.
How people rated its answers
We have not confirmed an Arena result for this exact version and thinking setting, so we leave Evidence Consensus blank. A result for another version or setting could be misleading.
API features to verify
The catalog lists tool-calling support and image input. These are declared features, not evaluations of tool accuracy or visual understanding.
Hosted access and official FP8/NVFP4 downloads are available. The October 29 weights use the Mistral Ecosystem Open-Weight License, with separate commercial licensing conditions for model-as-a-service businesses.
Mistral Large 4 API pricing
Input is what you send to the model; output is what it produces, including charged thinking tokens. Prices are per million tokens, the small pieces of text the model processes.
The October 30 native, OpenRouter and Gateway quotes are $0.70 input, $0.07 cached input and $2.10 output per million tokens. Historical quotes retain their original dates.
Direct/source price observed 2026-11-03; catalog offer rechecked 2026-11-03. Older observations remain dated until verified again.
At the baseline of 1,000 calls, 10,000 input tokens and 2,000 billed output tokens per call, the text API estimate is $11.2. Reasoning belongs in billed output. Caching, tools, retries, long-context premiums, service tiers, hosting and taxes are excluded.
Trace this model to the source
- LiveBench row ID
mistral-large-4-high- Question-set release
- 2026-06-25 · Source CSV ↗
- Catalog ID
- mistralai/mistral-large-4-0 ↗
- Arena configuration
- No verified matching observation
- Reviewed snapshot
- 2026-11-03
Compare alternatives
For production decisions, test examples from your own work using the same tools and a clear definition of success. Neither a large context window nor a strong benchmark score guarantees reliability in your own agent.