Mistral Large vs Mistral Large 4
Different efforts: non-thinking · thinking
Humanity's Last Exam Humanity's Last Exam: Very hard expert questions across many fields, written to stump today’s models
Mistral Large3.5
Mistral Large 435
Terminal-Bench 4.0 Terminal-Bench 4.0 · source setting 2: Agents use a terminal to complete executable coding, system administration and research tasks in the 4.0 suite
Mistral Large—
Mistral Large 426.8
AutomationBench-AA AutomationBench-AA: Agents orchestrate SaaS workflows through APIs; objectives receive credit only when the task has no guardrail violation
Mistral Large—
Mistral Large 459.9
Arena Text · creative writing Arena Text · creative writing: Human pairwise preferences in the stated text vote pool
Mistral Large1,276
Mistral Large 41,368
MMMU-Pro (vision) MMMU-Pro (vision): College-level questions that require understanding images: charts, diagrams, photos
Mistral Large—
Mistral Large 476.4
| Benchmark | Mistral Large | Mistral Large 4 |
|---|---|---|
| Humanity's Last Exam | 3.5 | 35 |
| Terminal-Bench 4.0 | — | 26.8 |
| AutomationBench-AA | — | 59.9 |
| Arena Text · creative writing | 1,276 | 1,368 |
| MMMU-Pro (vision) | — | 76.4 |
Specs
| Spec | Mistral Large | Mistral Large 4 |
|---|---|---|
| Input, per 1M | $2 | $1.36 |
| Output, per 1M | $6 | $4.18 |
| Cost per task? | — | $1.13 |
| Speed, tok/s | — | 67 |
| Context | 128K | 1M |
| Weights | Closed | Closed |
| Released | Feb 2024 | Oct 2026 |
| See images | — | ✓ |
| Read PDFs | Partial | ✓ |
| Hear audio | — | — |
| Watch video | — | — |
| Call tools | ✓ | ✓ |
| Force a tool call | Partial | Partial |
| Return strict JSON | ✓ | ✓ |
| Search the web | ? | ? |
| Run code | ? | ? |
| Use a computer | ? | ? |
What breaks when you switch
Mistral Large → Mistral Large 4
Infinite whitespace stream›
In JSON Object mode, include JSON in the system or user prompt to avoid a possible infinite whitespace stream.
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.
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.
10% upcharge · Mistral regional endpoints›
Regional inference charges 1.1× standard list prices for input, output, cached reads and cache writes.
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.
Mistral Large 4 → Mistral Large