Rerank is the cheap half of RAG, which is why a free rerank API matters

Everyone budgets RAG around embeddings, but embeddings are a one-time cost per document. Rerank runs on every query, and it's the step that actually fixes retrieval quality — vector similarity gets you candidates, rerank puts the right one first.

That asymmetry is why Cohere's free tier is interesting specifically for rerank. Embedding a corpus once on any provider is cheap. A rerank model you can call per-query at $0 changes what a hobby RAG app can afford to do on every request: over-retrieve 30 candidates, rerank, keep 5. That pattern usually beats a better embedding model, and it's free where it counts.

Trial keys are rate-capped rather than token-capped, which fits query-time use — the cap you feel is requests per minute, not corpus size.

Where the free tier's caps land and the embed-vs-rerank split in practice: https://toolfreebie.com/cohere-rag-api/