What is Codestral 25.08?
Codestral 25.08 is Mistral AI's code-completion model, released at the end of July 2025 and still generally available on Mistral's API. It specialises in fill-in-the-middle (FIM) completion: given the code before and after the cursor, it writes what goes between, quickly and many times a minute, which is what inline autocomplete needs. It also supports chat, function calling and structured outputs.
Mistral's coding line now has two halves: Codestral for fast, synchronous completion, and Devstral for multi-step agent work such as refactoring across files. The Kilo Code editor extension, among others, uses Codestral as its default autocomplete model.
On Mistral's API it costs $0.30 per million input tokens and $0.90 per million output tokens, half on batch, with a dedicated FIM endpoint. Mistral's docs list a 128K context, while OpenRouter lists 256K. Mistral also offers it for self-hosted enterprise deployment.
How much does Codestral 25.08 cost?
| Plan | Price | What you get |
|---|---|---|
| Mistral API | $0.30 in / $0.90 out per 1M tokens | FIM and chat endpoints |
| Batch | $0.15 in / $0.45 out per 1M tokens |
Mistral API: $0.30 input / $0.90 output per 1M tokens; batch $0.15 / $0.45. Enterprise self-hosting by arrangement.
Is Codestral 25.08 worth it?
Codestral does one job: fast inline completion. Mistral's numbers for the 25.08 update are relative rather than absolute, but they target what matters in an editor: 30% more completions accepted, 10% more suggested code kept, and half as many runaway generations that keep writing past where they should stop. The widely quoted 95.3% fill-in-the-middle score belongs to the earlier 25.01 version.
It is a specialist, and general rankings show it. It sits far down overall intelligence leaderboards, and Mistral itself steers multi-file and reasoning-heavy work to Devstral 2. That split is sensible: an autocomplete model needs low latency and low cost per call far more than deep reasoning.
On cost it is hard to argue with. At $0.30 in and $0.90 out per million tokens, a developer generating a few thousand completions a day costs cents. Tools including Kilo Code ship it as their default autocomplete, with a dedicated Codestral key for free use.
The question is age. Codestral 25.08 dates from July 2025, and newer small coding models, including Mistral's own Devstral Small 2 under Apache 2.0, can also complete code locally. Codestral's licence for self-hosting is commercial, not open. Teams that want open weights for local autocomplete should look at Devstral Small 2 or Qwen's coder models instead.
Where it earns its place
An IDE or editor-extension vendor
Offer inline autocomplete to users
Call the /v1/fim/completions endpoint with the code before and after the cursor. At $0.30/$0.90, a million completions of 1,000 input and 50 output tokens cost about $345.
A developer using Kilo Code
Get free autocomplete in VS Code
Kilo Code uses Codestral 25.08 as its default autocomplete; set a dedicated Codestral key, since a general Mistral key does not work for it.
An enterprise with strict code-privacy rules
Run autocomplete without sending code to a cloud API
Ask Mistral for a self-hosted Codestral deployment, or use the Apache-licensed Devstral Small 2 on your own GPUs instead.
Before you commit
- Setup
- Call codestral-2508 (or codestral-latest) on Mistral's API through the FIM or chat endpoints; editor plugins such as Kilo Code and Continue support it directly.
- Integrations
- Mistral FIM and chat APIs, batch API, OpenRouter; IDE extensions including Kilo Code.
- Data and privacy
- On Mistral's API, free-tier data may be used for training unless you opt out; pay-as-you-go users can opt out. Self-hosting is under a commercial licence by agreement.
- Cost at scale
- $0.30/$0.90 per million tokens. A 50-developer team making 2,000 completions each a day costs roughly $35 a day at 1,000 input and 50 output tokens per completion.
- Limits
- Specialist completion model: weak at reasoning and multi-file changes. Context 128K per Mistral's docs. No open-weight release of this version.
What's good
- Built for low-latency fill-in-the-middle completion
- $0.30/$0.90 per million tokens, half on batch
- Default autocomplete in tools such as Kilo Code
- Fewer runaway completions than earlier versions, per Mistral
What's not
- Released July 2025
- Weak on reasoning and agent tasks; Mistral points those to Devstral
- Self-hosting is a commercial licence, not open weights
- Context listed as 128K by Mistral and 256K by OpenRouter
Best for
Developers and IDE vendors who want cheap, fast inline code completion through an API.
Codestral 25.08 alternatives
Codestral 25.08: common questions
- Is Codestral 25.08 worth it?
- Still a sensible, cheap choice for inline autocomplete, but it is fifteen months old and the wrong tool for anything beyond completing the next few lines. We rate it 3.9 out of 5.
- How much does Codestral 25.08 cost?
- Mistral API: $0.30 input / $0.90 output per 1M tokens; batch $0.15 / $0.45. Enterprise self-hosting by arrangement.
- What are the downsides of Codestral 25.08?
- Released July 2025; weak on reasoning and agent tasks; Mistral points those to Devstral; self-hosting is a commercial licence, not open weights; context listed as 128K by Mistral and 256K by OpenRouter.
- Who is Codestral 25.08 best for?
- Developers and IDE vendors who want cheap, fast inline code completion through an API.
- How much does Codestral 25.08 cost for a team or in production?
- $0.30/$0.90 per million tokens. A 50-developer team making 2,000 completions each a day costs roughly $35 a day at 1,000 input and 50 output tokens per completion.
- Is Codestral 25.08 safe for company data?
- On Mistral's API, free-tier data may be used for training unless you opt out; pay-as-you-go users can opt out. Self-hosting is under a commercial licence by agreement.
- What are the best Codestral 25.08 alternatives?
- The closest alternatives we have reviewed are Devstral Small 2, Devstral 2 and Qwen3.8-27B.
Answers checked Oct 25, 2026.