Embedding models

OpenRouter keeps an embeddings ranking beside the chat catalog. Chat slugs such as DeepSeek and Sonnet are the wrong tool for “best embedding model.” Embeddings turn text into vectors for search. You care about recall on your documents, the vector size you must store, and the price per million input tokens.

What to compare

• Recall on a sample of your own queries, not a generic leaderboard alone.
• Dimension. A 3072-d index costs more to store than a 768-d index.
• Input price. Embedding calls are almost all prompt tokens.
• Context. Long chunks that get truncated will not match the leaderboard.
• Language. A model trained mostly on English will miss a Chinese or mixed corpus.

The rankings page states that OpenRouter’s ranking datasets are under CC BY 4.0. Quote the table with attribution if you republish it. The order moves. Pin the model id in your indexer so a later “top model” does not silently rebuild the index in a different space.

Do not mix vectors from two models in one index. Switching models means re-embedding.