Mistral Releases Codestral Embed, Code-Specialized Embedding Model at $0.15 Per Million Tokens

Mistral AI has released Codestral Embed, its first embedding model specialized for code retrieval and semantic understanding. The model is available via API under the identifier codestral-embed-2505 at $0.15 per million tokens, with a 50% discount available through Mistral's batch API.

Technical Specifications

Codestral Embed supports an 8192-token context window and offers flexible output configurations. The model can generate embeddings at different dimensions and precisions, with dimensions ordered by relevance. According to Mistral, even at reduced dimension 256 with int8 precision, the model outperforms competing code embedding models.

For retrieval use cases, Mistral recommends chunking datasets into 3000-character segments with 1000-character overlap rather than using the full context window, as larger chunks reportedly degrade retrieval performance.

Benchmark Performance

Mistral claims Codestral Embed outperforms Voyage Code 3, Cohere Embed v4.0, and OpenAI's large embedding model across multiple code retrieval benchmarks. The company highlights particularly strong performance on:

• SWE-Bench Lite: Real-world GitHub issue resolution, identifying files requiring modification
• CodeSearchNet: Code-to-code and documentation-to-code retrieval from GitHub repositories
• Text2SQL tasks: Spider, WikiSQL, and synthetic SQL generation benchmarks
• Algorithm challenges: APPS, CodeChef, MBPP+, and competitive programming problems
• Data science: DS-1000 matching questions to implementations

Mistral evaluated the model across categories including SWE-Bench, code-to-code retrieval, Text2Code from GitHub data, Text2SQL, algorithmic problems, and data science tasks. Specific numerical scores were not disclosed in the announcement.

Use Cases

The company positions Codestral Embed for four primary applications:

1. Retrieval-augmented generation: Context retrieval for code completion, editing, and explanation in coding assistants and agent frameworks
2. Semantic code search: Natural language or code-based queries across documentation and development tools
3. Duplicate detection: Identifying functionally similar code segments for reuse optimization and license enforcement
4. Code analytics: Unsupervised clustering for repository analysis and architecture pattern identification

Availability

Codestral Embed is accessible through Mistral's API and batch API. On-premises deployments require direct contact with Mistral's applied AI team. The company has published documentation and a cookbook with examples for code agent retrieval implementations.

What This Means

Codestral Embed represents Mistral's first specialized product for developer tooling, competing directly with established code embedding models from Voyage AI, Cohere, and OpenAI. The $0.15 per million token pricing positions it competitively in the embedding market, though direct cost comparisons depend on competitors' pricing for comparable dimensions and precision levels. The model's flexible dimension reduction feature could provide cost savings for applications where storage efficiency matters more than maximum retrieval accuracy. Its strong claimed performance on SWE-Bench Lite, a benchmark using real GitHub issues, suggests potential utility for coding agents and automated software engineering tools.