Mistral's specialized coding model. Optimized for code generation, completion, and analysis.
Codestral is Mistral AI's specialized coding model family, first released May 29, 2024 as a 22.2B-parameter open-weight model. It was designed from the ground up for code generation, completion, and fill-in-the-middle (FIM) tasks rather than general-purpose chat. Over its lifecycle, Codestral has seen two major revisions: Codestral 25.01 (Jan 2025) expanded the context window to 256K tokens and introduced a more efficient tokenizer, and Codestral 25.08 (Jul 2025) delivered roughly 30% more accepted code suggestions. The model supports 80+ programming languages.
For TheRouter operators, Codestral occupies a distinct niche: it is a fast, lightweight code completion model rather than an agentic coding assistant. Its native Fill-in-the-Middle (FIM) capability makes it ideal for IDE autocomplete, short single-file scaffolding, and test generation — workloads where latency matters more than deep reasoning. It is not designed for multi-file refactoring, complex agent loops, or tool-calling workflows; those are better served by Mistral's Devstral 2 or Claude Opus 4.7. On TheRouter, Codestral is reachable through the standard OpenAI-compatible chat completions endpoint, and its $0.30 / $0.90 per million tokens pricing makes it one of the most cost-effective code-completion options available through any router.
| Type | Rate |
|---|---|
| Input | $0.324 / 1M tokens |
| Output | $0.972 / 1M tokens |
| Release date | 2024-05-29 (original); 2025-01-13 (25.01); 2025-07-30 (25.08)mistral.ai ↗ | verified |
| Architecture | 22.2B-parameter dense decoder-only transformer (original). 25.01/25.08 architecture not publicly disclosed — estimated below 100B parameters with improved tokenizerhuggingface.co ↗ | to verify |
| Context window | 32K (original); 256K (25.01+); TheRouter entry shows 256Kmistral.ai ↗ | verified |
| Native fill-in-the-middle (FIM) | Yes — Codestral's core design feature. Natively trained for FIM completion; 25.01 improved FIM quality with new tokenizermistral.ai ↗ | verified |
| Languages | 80+ programming languages including Python, Java, JavaScript, TypeScript, C++, Go, Rust, SQLmistral.ai ↗ | verified |
| License | Mistral AI Non-Production License (original 22B open weights). Codestral 25.01+ is API-only — weights not publicly releasedhuggingface.co ↗ | verified |
| Successor | Mistral Devstral 2 (May 2025) is the agentic coding successor with tool use and multi-file capabilities. Codestral 25.08 remains available for pure code completion workloads | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
HumanEval pass@1 Original Codestral 22B (May 2024); Python single-function code generation | 81.1%% | mistral.ai ↗ | |
HumanEval pass@1 Codestral 25.01 (Jan 2025) — Python single-function pass@1 improvement over original | 86.6%% | mistral.ai ↗ | |
HumanEval FIM average Codestral 25.01 fill-in-the-middle across Python, Java, JavaScript | 85.9%% | mistral.ai ↗ | |
MBPP sanitized pass@1 Codestral 25.01 — Python program synthesis | 80.2%% | mistral.ai ↗ | |
CruxEval Python output prediction — original Codestral 22B (May 2024) | 75.5%% | mistral.ai ↗ | |
RepoBench EM Long-range repository-level code completion — original Codestral 22B; outperformed Code Llama 70B and DeepSeek Coder 33B at launch | 34.0%% | mistral.ai ↗ | |
Spider SQL Text-to-SQL generation — Codestral 25.01 | 66.5%% | overchat.ai ↗ |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "mistral/codestral",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI-compatible chat completion. Codestral responds best to explicit code-generation prompts.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "mistral/codestral",
"messages": [
{"role": "user", "content": "Write a Python function to merge two sorted arrays into one sorted array."}
],
"temperature": 0.2
}'Mistral released Codestral 25.08 alongside Codestral Embed, Devstral 2, and the Mistral Vibe CLI — forming an end-to-end enterprise coding stack. Codestral 25.08 delivers roughly 30% more accepted code suggestions over 25.01, with improved latency.
Mistral released Codestral 25.01 with a 256K context window (up from 32K), a more efficient tokenizer, and roughly 2x faster code generation. HumanEval pass@1 improved from 81.1% to 86.6%. FIM scores reached 85.9% across Python, Java, and JavaScript.
No — Codestral is specialized for code generation and completion. For general chat, routing to Mistral Small 4 or Mistral Large 3 will give better results at comparable or lower cost. The standard-models.yaml entry lists tools/tool_choice in supported_params, but in practice Codestral's design emphasis is on raw code completion, not agentic tool use.
FIM is a code completion technique where the model fills in a missing section of code given surrounding context (prefix and suffix). Codestral is natively trained for this, making it ideal for IDE autocomplete where the cursor is in the middle of a file. Codestral 25.01 scored 85.9% average on HumanEval FIM across Python, Java, and JavaScript.
Only the original Codestral 22B (May 2024) has publicly available open weights, licensed under Mistral AI's Non-Production License. Codestral 25.01 and 25.08 are API-only — their weights have not been publicly released. Mistral does offer fine-tuning for Codestral through their own API platform, so custom training is possible via that route.
Codestral is a focused code completion model — fast, cheap, and specialized for single-file generation and inline FIM. Devstral 2 (May 2025) is the agentic successor built on top of the Mistral coding stack, adding tool calling, autonomous debugging, multi-file refactoring, and MCP integration. If you need IDE autocomplete or cheap code generation, use Codestral. If you need an agent that edits whole repositories, use Devstral 2 via TheRouter.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | mistral.ai ↗ | 2026-05-27 | verified |
| Architecture | huggingface.co ↗ | 2026-05-27 | to verify |
| Context window | mistral.ai ↗ | 2026-05-27 | verified |
| Native fill-in-the-middle (FIM) | mistral.ai ↗ | 2026-05-27 | verified |
| Languages | mistral.ai ↗ | 2026-05-27 | verified |
| License | huggingface.co ↗ | 2026-05-27 | verified |
| Successor | — | — | verified |
| HumanEval pass@1 | mistral.ai ↗ | 2026-05-27 | verified |
| HumanEval pass@1 | mistral.ai ↗ | 2026-05-27 | verified |
| HumanEval FIM average | mistral.ai ↗ | 2026-05-27 | verified |
| MBPP sanitized pass@1 | mistral.ai ↗ | 2026-05-27 | verified |
| CruxEval | mistral.ai ↗ | 2026-05-27 | verified |
| RepoBench EM | mistral.ai ↗ | 2026-05-27 | verified |
| Spider SQL | overchat.ai ↗ | 2026-05-27 | to verify |
| Codestral 25.08: +30% accepted suggestions, complete Mistral coding stack | mistral.ai ↗ | 2026-05-27 | verified |
| Codestral 25.01: 256K context, 2x speed, improved FIM | mistral.ai ↗ | 2026-05-27 | verified |
| Is Codestral a good general-purpose chatbot? | mistral.ai ↗ | 2026-05-27 | to verify |
| What is fill-in-the-middle (FIM) and why does it matter? | mistral.ai ↗ | 2026-05-27 | to verify |
| Can I fine-tune Codestral on my own codebase? | huggingface.co ↗ | 2026-05-27 | to verify |
| How does Codestral compare to Devstral 2? | mistral.ai ↗ | 2026-05-27 | to verify |