OpenRouter publishes 2026 guide to choosing embedding models for various use cases

On September 11, 2026, OpenRouter released a guide evaluating 19 embedding models, covering 37 entries, to help users select the best options for different scenarios. The guide highlights models like openai/text-embedding-3-small for English RAG, which offers 8192 token context and adjustable dimensions, and voyage-4-large for handling larger inputs. For multilingual tasks, qwen3-embedding-8b is recommended, supporting over 100 languages. Code search benefits from voyage-code-4, while text-image retrieval uses google/gemini-embedding-2 or voyage-multimodal-3.5. OpenRouter tested these models using 28 checks, noting that 16 paid models returned vectors for all inputs, while free routes faced 404 errors due to privacy settings. The guide emphasizes the importance of testing and compatibility, as vectors from different model families do not share the same coordinate space. Developers are advised to store model details with index metadata to ensure future verification.