Felix Laurent
Cohere introduced Context Focus on oct 21, an opt-in feature that returns precise passages with relevance scores instead of whole documents. Cohere says it reduces token consumption by up to 41% while maintaining accuracy.
what cohere announced
The announcement post is a follow-on to a larger launch. The table lists what the sources state.
| item | what the sources say |
|---|---|
| feature | Context Focus, described as brand-new |
| output | precise passages with relevance scores |
| token claim | up to 41% lower consumption, per Cohere |
| mode | opt-in, works like a dial |
| product | Rerank 5, in Pro and Fast |
The feature is part of Rerank 5, not a standalone product.
how context focus works
Cohere's docs say both Rerank 5 versions support Context Focus. It returns the most relevant passages within each document in the same call. Calibrated relevance scores come with them.
Cohere says the feature works like a dial. The user controls how much context and quality to keep. Because it is opt-in, existing setups are unchanged until it is switched on.
what the 41% figure does not say
The 41% is an up-to figure from Cohere's own statement. The sources give no benchmark, baseline or definition of maintaining accuracy.
The sources also give no pricing for the feature. A team with a retrieval pipeline would need to test it on its own documents to see the real saving.
Until Cohere publishes the method, the number is a best case.
