Codestral

Available standard

Purpose-built for code generation and software engineering

Released May 29, 2024

API Documentation View Provider

Context Window: 32K
TTFT: N/A
Speed: N/A

Max Output: 8K
Training Cutoff: 2024-04

Last tested: 2026-01-15

About

Codestral is Mistral's dedicated code model, trained specifically for software development tasks. With a 32K context window optimized for file-level understanding, it excels at code completion, generation, refactoring, and documentation across 80+ programming languages. The go-to choice for IDE integrations and developer tools.

Capabilities

code-completion code-generation streaming fill-in-middle low-latency refactoring

Pricing

Input: $0.20/M
Output: $0.60/M

Live pricing Static pricing

Details

Codestral

Codestral is Mistral’s answer to the specialized code model category. Rather than retrofitting a general-purpose model for code, Mistral trained Codestral from the ground up on software engineering data, resulting in a model that truly understands how developers work.

Optimized for Development

80+ Languages: From mainstream languages like Python, JavaScript, TypeScript, Java, C++, and Go to specialized languages like Rust, Kotlin, Swift, and even legacy languages like COBOL and Fortran.

Fill-in-the-Middle: Native support for code infilling, essential for IDE integration. Codestral can complete code given surrounding context, not just generate from a prefix.

Ultra-Low Latency: At 35ms average response time and 200 tokens per second, Codestral feels instantaneous in real-time coding environments.

IDE Integrations

Codestral powers coding assistants across the ecosystem:

VS Code: Via Continue, Cody, and other extensions
JetBrains: Native plugin support
Vim/Neovim: Through LSP integrations
Jupyter: Interactive notebook assistance

Developer Workflows

Code Generation: Describe what you need in natural language, get working code.

"Write a Python function that validates email addresses using regex"

Refactoring: Improve existing code structure and performance.

"Refactor this function to use async/await instead of callbacks"

Documentation: Generate docstrings, comments, and README content.

"Add comprehensive docstrings to this class following Google style"

Test Generation: Create unit tests from implementation code.

"Generate pytest tests for this authentication module"

Pricing for Scale

At $0.20/M input tokens, Codestral is economical enough for high-frequency use cases like real-time code completion. A typical developer session generates thousands of requests; Codestral keeps costs manageable.

European Engineering

Built in Paris by Mistral’s team of former DeepMind and Meta researchers, Codestral represents European excellence in AI for developers. It offers a compelling alternative to American code models with transparent pricing and data handling.