Codestral 25.08

Mistral AI announced Codestral 25.08 on July 30, 2025 as the centerpiece of its enterprise "Mistral Coding Stack," bundling the model with Codestral Embed, Devstral agents, and the Mistral Code IDE extension for VS Code and JetBrains. The release targets enterprise pain points such as SaaS-only deployment, lack of model access, fragmented agents, and absent observability. Codestral is positioned for cloud, VPC, on-prem, and air-gapped environments to meet regulated-industry requirements. The announcement frames Codestral 25.08 as a low-latency code-generation model optimized for high-frequency IDE tasks, including code completion, fill-in-the-middle, code correction, and test generation across 80+ languages. Mistral promotes the stack as cutting development, review, and testing time by 50%. The post also highlights customization needs, such as fine-tuning to internal codebases via post-training workflows and extensibility beyond closed copilots.

The AI/TLDR catalog confirms Codestral 25.08 (API id codestral-2508) was released July 30, 2025 as Mistral's flagship code-generation model within the Mistral Coding Stack, alongside Devstral agents, the Codestral Embed retrieval model, and the Mistral Code IDE plugin for VS Code and JetBrains. It is a proprietary Premier model accessible via the Mistral API and resellers such as OpenRouter, with no downloadable weights, though enterprise cloud, VPC, and on-prem deployments are offered. Mistral-reported improvements over Codestral 25.01 include 30% more accepted completions, around 10% more retained code after acceptance, 50% fewer runaway generations, and roughly 5% gains on instruction-following and average MultiplE code ability. The model handles code completion, fill-in-the-middle, code correction, and test generation across 80+ languages at $0.30 per million input tokens and $0.90 per million output tokens.

Techzine reports that Codestral 25.08 delivers 30% more accepted code suggestions, retains 10% more code after edits, and halves runaway generations versus its predecessor. The model supports fill-in-the-middle completion and is tuned for low-latency production environments, and it can run unchanged across cloud, VPC, and on-premises infrastructure. The full Mistral coding stack launched the same day via console.mistral.ai and enterprise request channels. Alongside Codestral 25.08, the stack bundles agentic Devstral workflows that score 53.6% (Small) to 61.6% (Medium) on SWE-Bench Verified, outperforming Claude 3.5 and GPT-4.1-mini per Mistral. Mistral Code integrates inline completion, one-click commit messages, and semantic search into JetBrains and VS Code. Mistral also claims its codebase search component outperforms OpenAI and Cohere on retrieval quality.

The official Mistral documentation lists Codestral 25.08 (API id codestral-2508) as a Premier tier "GAPremierv25.08" coding model released July 30, 2025, specializing in low-latency fill-in-the-middle and code generation. It exposes a 128K-token context window with text input/output and is priced at $0.30 per million input tokens and $0.90 per million output tokens. The page is dated and confirms general availability. Supported endpoints include /v1/chat/completions, /v1/fim/completions, /v1/batch, and /v1/conversations, with features such as Structured Outputs, Function Calling, Predicted Outputs, Document QnA, Prefix, and Batching. The model is documented for IDE-integrated completion and instruction-following coding workflows. No image or audio modalities are listed, confirming text-only operation.