Codestral 25.08 Review: Benchmarks, Pricing & IDE Setup
Mistral's dedicated code completion model gets its August 2025 update — improved fill-in-the-middle accuracy, 256k context, 80+ language support, and $0.3/M input pricing. Here's what changed and whether it belongs in your developer toolchain.
Reviewed 2026-06-11 · Source: Mistral announcement
Quick Verdict
Codestral 25.08 is the best dedicated code completion API for teams who want low-latency FIM without paying for a general-purpose frontier model. At $0.3/M input, it's significantly cheaper than Claude or GPT-4o for the same completions. The 256k context window eliminates file-chunking headaches, and the 80+ language support means one API covers polyglot projects. The main trade-off: it's proprietary (no self-hosting), and it's not an agent — for multi-step coding tasks, Devstral is the right tool. For inline IDE completion and code generation, Codestral 25.08 is a strong choice.
What's New in Codestral 25.08
Improved FIM accuracy
Codestral 25.08 refines fill-in-the-middle performance over the January 2025 release. FIM lets the model complete code given both a prefix and a suffix — the key capability for inline IDE completions where you're editing in the middle of a file. The 25.08 update improves precision across multi-language FIM tasks, reducing completions that don't fit syntactically or semantically with surrounding code.
256k token context window
The 256k context window lets Codestral hold an entire large file — or multiple related files — in a single call. For code completion and test generation, this eliminates most chunking strategies you'd otherwise need. You can pass a full 10,000-line codebase and generate targeted completions that reference any function, class, or variable defined anywhere in the file.
Broader language coverage
Codestral 25.08 supports 80+ programming languages. Core coverage includes Python, JavaScript, TypeScript, Go, Rust, Java, C, C++, SQL, Shell, and PHP. Extended coverage includes Kotlin, Swift, R, Lua, Haskell, Elixir, Erlang, COBOL, and many more. For polyglot teams, this means a single model API handles completions across different stacks without needing separate fine-tuned models per language.
Low-latency inference design
Unlike general-purpose frontier models that maximize reasoning depth, Codestral's architecture prioritizes throughput and latency. Mistral has optimized it for high-frequency, short-duration calls — the pattern of keystroke-level IDE completion. This makes Codestral meaningfully faster per completion than using a general model like Mistral Large or Claude Sonnet for the same task.
IDE and editor integrations
Codestral 25.08 has official integrations with VS Code (via the Continue.dev extension), JetBrains IDEs, Jupyter notebooks, neovim, and Emacs. The Continue.dev plugin is the primary path for VS Code — it supports both chat and inline FIM completion. All integrations use the same La Plateforme API key, so there's no separate authentication or account setup.
Key Specs
Core capabilities and pricing for Codestral 25.08.
| Spec | Codestral 25.08 | Notes |
|---|---|---|
| Context window | 256k tokens | Full large-codebase and multi-file support |
| Language support | 80+ languages | Python, JS/TS, Go, Rust, Java, C/C++, and more |
| API pricing (input) | $0.3 / 1M tokens | Cheaper than most general-purpose frontier models |
| API pricing (output) | $0.9 / 1M tokens | Competitive for high-volume completion workloads |
| Fill-in-the-middle (FIM) | Native support | First-class FIM — not a general model adaptation |
| Model version | 25.08 (Aug 2025) | Succeeds Codestral 25.01 (Jan 2025) |
Pricing
$0.3 / $0.9 per 1M tokens
Input $0.3/M · Output $0.9/M. Model ID: codestral-latest or codestral-25-08. Standard Mistral API account. Pay-as-you-go with no minimum commitment.
Free (limited)
Mistral offers limited free access to Codestral for personal/non-commercial use via codestral.mistral.ai. Rate-limited. Not suitable for production workloads.
Who Should Use Codestral 25.08
Codestral vs Alternatives
How Codestral 25.08 compares to other code completion options in 2026.
vs GitHub Copilot
Copilot is a managed product with billing, IDE plugins, and enterprise controls bundled in. Codestral is a raw API — more flexible and cheaper per token, but you own the integration. For individual developers, Copilot is easier. For teams building internal tools or needing API-level control, Codestral wins on cost and flexibility.
vs Devstral 2 (Mistral)
Devstral 2 is for agentic coding — planning, executing, and iterating on multi-step software tasks. Codestral is for completion — fast, single-turn fills. They serve different use cases: Codestral for IDE inline completion, Devstral for AI coding agents that can autonomously write and test features.
vs Claude / GPT-4o for code
General-purpose frontier models are excellent at code but not built for FIM or real-time completion. They're also 5–10× more expensive per token. For one-off code generation in a chat interface, Claude/GPT-4o are fine. For high-frequency inline completion or batch code generation at volume, Codestral's cost and latency profile is much more practical.
FAQ
What is Codestral 25.08?
Codestral 25.08 is Mistral AI's dedicated code completion model, released in August 2025. It's optimized for fill-in-the-middle (FIM) code completion, code correction, and test generation across 80+ programming languages. Unlike general-purpose LLMs, Codestral is designed specifically for low-latency, high-frequency IDE integration.
How does Codestral compare to GitHub Copilot?
Codestral is a direct API — you bring your own integration (VS Code, JetBrains, etc.) rather than using a managed subscription product. For teams who want model-level control, cost transparency, or to avoid Microsoft's data policies, Codestral offers an alternative. GitHub Copilot bundles more out-of-the-box UX; Codestral offers lower cost per token and more flexibility for custom pipelines.
What is fill-in-the-middle (FIM) and why does it matter?
FIM is the ability to complete code given both what comes before (prefix) and after (suffix) the cursor position. It's the technique that makes inline code completion feel smart — the model knows what comes next and what came before, so it doesn't generate a completion that breaks the surrounding code. Codestral has first-class FIM support; most general chat models don't.
Is Codestral open source?
No. Codestral 25.08 is a Premier (proprietary) model — weights are not publicly available. It's accessible via Mistral's La Plateforme API only. This is different from Mistral's open models like Mistral Small 4 (Apache 2.0) or Devstral 2 (open weights). If you need a self-hosted code model, Devstral 2 is Mistral's open-weights option for coding agents.
What's the difference between Codestral and Devstral?
Codestral is optimized for low-latency inline code completion and FIM — the keystroke-level autocomplete use case. Devstral is optimized for agentic software engineering — multi-step tasks like writing, debugging, and refactoring entire features. Use Codestral for IDE completion; use Devstral when you want an AI agent that can plan and execute multiple coding steps autonomously.
How do I integrate Codestral with VS Code?
Install the Continue.dev extension in VS Code, then configure it to use the Mistral API with the codestral-latest model. Continue.dev handles both inline tab-completion (FIM) and the chat sidebar. You'll need a La Plateforme API key from mistral.ai. The setup takes under five minutes.
Try Codestral 25.08
Get API access via Mistral La Plateforme. $0.3/M input · $0.9/M output.
View Codestral on MistralAffiliate disclosure: Some links on this page are affiliate links. If you sign up through them, AISO Tools may earn a commission at no extra cost to you. This never affects our rankings or reviews.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.
Join thousands of professionals who discover the best AI tools every week. No spam — unsubscribe anytime.