Codestral 2508
Neural Network
Codestral 2508 for code by Mistral: low latency, fill-in-the-middle completion, bug fixing, and test generation.
Max answer length
(in tokens)
Context size
(in tokens)
Prompt cost
(per 1M tokens)
Answer cost
(per 1M tokens)
How it works Codestral 2508?
Codestral 2508 is a Mistral language model released in late July 2025, optimized for coding tasks. It specializes in low-latency, high-frequency tasks: fill-in-the-middle code completion, bug fixing in existing snippets, and test generation—everything repeated dozens of times an hour in an editor. In BotHub, Codestral is available without a VPN or foreign card; payment is made in rubles via Russian cards based on actual token usage, and unused tokens do not expire. Connection to an editor or CI is via the unified BotHub API: you can switch projects to another model without rewriting the integration, as over 250 models are gathered in one window. Chats are encrypted via AES-GCM, and request content is not saved—crucial when proprietary source code is in the prompt. Useful if you are tweaking autocompletion directly in the IDE, fixing bugs via stack traces, covering modules with unit tests, refactoring legacy code to team standards, or building an internal developer assistant.Frequently asked questions about Codestral 2508
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
This is a Mistral model focused on coding: writing functions, refactoring, explaining unfamiliar snippets, finding bugs, and autocompletion suggestions. A 256,000-token context allows loading large files and entire documentation, while the batch mode is convenient for batch processing tasks.
The output limit is 204,800 tokens per response. This is enough to generate an entire module, rewrite a large file with tests, or provide a detailed code analysis without breaking the task into many separate requests.
A separate step-by-step reasoning mode is not noted in our data, so we won't promise it. In practice, it helps to formulate the request in steps and ask the model to first describe a solution plan, then provide the code. For complex reasoning chains, it is easy to switch to another model in BotHub.
Yes, the model supports function calling and structured JSON output. This allows connecting it to external tools, databases, and custom services, obtaining a predictable response format, and embedding the model into agent scenarios via the unified OpenAI-compatible BotHub API.
Besides text, the model accepts documents, including native formats, which is convenient for parsing specifications, technical requirements, and code dumps. Images, audio, and video input are not supported; for those, BotHub has separate multimodal models in the same window.
Codestral was created primarily for programming languages and coding tasks. There is no information in our data regarding prompt and response languages, so it is safer to test the model with your own tasks: in BotHub, this can be done without a VPN or foreign card, with payment in rubles.