Overview
Codestral 25.01 (API id codestral-2501) is a code-generation model that Mistral AI released on January 13, 2025 as an update to its original Codestral line. It is tuned for low-latency, high-frequency developer tasks such as code completion, fill-in-the-middle (FIM), code correction, and test generation across more than 80 programming languages, and it shipped to developers through IDE plugin partners (notably Continue for VS Code and JetBrains).
The headline change in Codestral 25.01 is a more efficient architecture and improved tokenizer that Mistral says generates and completes code about 2x faster than the first Codestral, paired with a much larger 256K-token context window (up from 32K). On launch it debuted at #1 on the Copilot Arena leaderboard (joint top alongside Claude 3.5 Sonnet and DeepSeek V2.5 FIM) and scored 86.6% on HumanEval for Python. It was benchmarked against Codestral 24.05 (22B), Code Llama 70B, DeepSeek Coder 33B / V2 Lite, and OpenAI's FIM API.
Unlike the original open-weight Codestral 22B (24.05), Codestral 25.01 was never released as downloadable weights — it was API-only, available via Mistral's la Plateforme (codestral-latest), Google Cloud Vertex AI, Azure AI Foundry, and Amazon Bedrock, with local/VPC deployment offered only for enterprise residency needs. Mistral has since deprecated it (November 6, 2025) and retired it (November 30, 2025), pointing users to the newer Codestral (25.08).
| Released | 2025-01-13 |
|---|---|
| License | Proprietary (Mistral commercial / API-only — not released as open weights, unlike the original Codestral 22B) |
| Weights | API only |
| Parameters | Not publicly disclosed |
| Context | 256K |
| Max output | Not disclosed |
| Architecture | Decoder-only Transformer tuned for low-latency code completion and fill-in-the-middle (FIM). Mistral describes 25.01 as having a more efficient architecture and an improved tokenizer than the original Codestral, generating and completing code roughly 2x faster. Parameter count not publicly disclosed. |
| Knowledge cutoff | Not disclosed |
| Modalities | Text |
| Status | Retired (deprecated 2025-11-06, retired 2025-11-30; superseded by Codestral 25.08). Mistral routes users to the newer 'Codestral' model. |
Benchmarks
- HumanEval (Python, pass@1)86.6%
- MBPP80.2%
- CruxEval55.5%
- LiveCodeBench37.9%
- RepoBench38%
- Spider (text-to-SQL)66.5%
- CanItEdit50.5%
- FIM pass@1 (average of Python, Java, JS)95.3%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $0.30 / 1M tokens per 1M tokens |
|---|---|
| Output | $0.90 / 1M tokens per 1M tokens |
Mistral API list price for the 'Codestral' endpoint (which codestral-2501 served under); the same rates carry on the current Codestral model. Mistral did not publish a separate price specific to the 25.01 build.
Strengths
- Low-latency code completion and fill-in-the-middle (FIM), tuned for real-time IDE autocomplete
- About 2x faster code generation than the original Codestral, via a more efficient architecture and improved tokenizer
- Large 256K-token context window for reasoning over big files and multi-file repositories
- Broad coverage of 80+ programming languages, plus code correction and test generation
- Strong code benchmarks at launch — 86.6% HumanEval (Python) and a #1 debut on the Copilot Arena leaderboard
Best for
- Real-time IDE code completion and fill-in-the-middle inside editors like VS Code and JetBrains (via the Continue plugin)
- Code generation, refactoring, and bug fixing across many programming languages
- Automated test generation and code correction
- On-prem or VPC coding assistants for enterprises with data- and model-residency requirements
How to access
| Provider | Model ID |
|---|---|
| Mistral AI ↗ | codestral-2501 |
| OpenRouter ↗ | mistralai/codestral-2501 |
Codestral — every version
The full lineage of the Codestral line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Codestral (25.08)current | 2025-07-30 | — | Open weights |
| Codestral 25.01 | 2025-01 | — | Open weights |
| Codestral 22B (24.05) | 2024-05-29 | — | Open weights |
FAQ
What is Codestral 25.01?
Codestral 25.01 (API id codestral-2501) is a code-generation model Mistral AI released on January 13, 2025. It is tuned for low-latency tasks like code completion, fill-in-the-middle (FIM), code correction, and test generation across 80+ programming languages, with a more efficient architecture and improved tokenizer that generate code about 2x faster than the original Codestral, plus a 256K-token context window.
Is Codestral 25.01 open weights?
No. Unlike the original open-weight Codestral 22B (24.05), Codestral 25.01 was never released as downloadable weights. It was API-only — available through Mistral's la Plateforme, Google Cloud Vertex AI, Azure AI Foundry, and Amazon Bedrock — with local/VPC deployment offered only to enterprises with data-residency needs.
Is Codestral 25.01 still available?
No. Mistral deprecated codestral-2501 on November 6, 2025 and retired it on November 30, 2025, directing users to the newer Codestral model (25.08, codestral-2508). The codestral-latest alias now points to the current Codestral.
How much did Codestral 25.01 cost?
It was billed under Mistral's Codestral endpoint at $0.30 per million input tokens and $0.90 per million output tokens — the same rates Mistral still lists for the current Codestral model. Mistral did not publish a separate price for the 25.01 build.