Codestral
Last updated: April 22, 2026
Specs
| Vendor | Mistral AI |
|---|
| Released | January 13, 2025 |
|---|
| Context window | 256K tokens |
|---|
| Input | $0.3 / M tokens |
|---|
| Output | $0.9 / M tokens |
|---|
Benchmarks
| mmlu | 73.2 |
|---|
| gpqa | 41 |
|---|
| humanEval | 89.7 |
|---|
| arenaElo | 1254 |
|---|
| sweBench | 35 |
|---|
| sweBenchPro | 26 |
|---|
| liveCodeBench | 55 |
|---|
| aiderPolyglot | 51 |
|---|
| terminalBench | 18 |
|---|
| mmluPro | 50 |
|---|
| hle | 3 |
|---|
| aime | 30 |
|---|
| math500 | 65 |
|---|
Strengths
- Purpose-built for coding, strong HumanEval for its price
- Supports 80+ programming languages
- FIM (fill-in-the-middle) for IDE use cases
Weaknesses
- Weak outside code tasks (low GPQA)
- No vision or agentic capabilities
- Not the best for code review reasoning
Best for
- Custom IDE autocomplete and inline suggestions
- Low-latency coding copilots
- Code translation and refactoring
Compare with other AI models