Cohere’s newest retrieval model comes in two speeds, and the trick is that they live in the same mathematical space.
Embed 5, announced this week, splits into a Pro tier tuned for retrieval quality and a Fast tier tuned for latency and cost on the query path. A team can index documents with Pro and answer live searches with Fast, as long as both sides output the same dimension. The family handles text, images or a fused combination, spans more than 100 languages and reads up to 128,000 tokens at a time.
Both tiers are available now on the Cohere API and Model Vault, Microsoft Foundry and Amazon SageMaker, with private deployments served through vLLM. Pricing runs $0.12 per 1M text tokens for Pro and $0.08 for Fast, with image inputs at $0.40 for either.
Cohere’s own numbers put a Pro index queried with Fast at 98.4, against 100 for Pro queried with Pro and 96.6 for an all-Fast setup. On the ViDoRe V3 benchmark Pro averaged 85.8, ahead of Voyage 4 Large at 83.7 and Gemini Embedding 2 at 83.2. Fast handled 377 documents per second to Pro’s 159.
One caveat: most scores use RCP-nDCG@10, a newer Cohere metric that leans on reranking, and outside replication has not yet happened.
