August 2026 Summaries

Voyage has introduced voyage-code-4, a code embedding model designed for coding agents that need accurate, low-latency semantic retrieval when investigating broad bug reports or repository-wide behaviors. The model was trained on a new corpus derived from issue-fixing pull requests, pairing natural-language problem descriptions with code changed in completed fixes to better identify code relevant to symptoms rather than named identifiers. On a 19-benchmark agentic retrieval suite, Voyage reports that voyage-code-4 exceeds voyage-code-3, Cohere Embed v4, Gemini Embedding 2, and OpenAI v3 large by 27.54%, 28.25%, 31.03%, and 48.58%, respectively, using NDCG@10; it also reports improvements across 28 traditional code-retrieval datasets. It supports embedding sizes from 256 to 2,048 dimensions through Matryoshka learning and float32, int8, and binary quantization options, while costing $0.12 per million tokens, one-third less than voyage-code-3. The model is available through the Voyage API and MongoDB Atlas Embedding and Reranking API, with the first 200 million tokens offered free.