Cohere launches Rerank 5 Pro and Fast with passage-level Context Focus and calibrated relevance scores

• Models: rerank-v5.0-pro and rerank-v5.0-fast (AIToolsReview).
• Context window: 32,768 tokens for query plus document. It handles multilingual text and semi-structured JSON (LLM Reference).
• Pricing: Pro costs $0.05 and Fast $0.02 per million tokens (Cohere).
• Speed: Fast processes about 187 documents per second on an A100, roughly 2.9× Pro's throughput (Cohere).
• Deployment: API, Microsoft Foundry, SageMaker, or private VPC and on-premises. It plugs into existing keyword, vector or hybrid search without re-indexing (Cohere).
• Model size: reported as about 2B parameters for Pro and about 500M for Fast in a community PR (GitHub PR). The Cohere launch post does not state these figures.

• Context Focus: The same rerank call returns scored character spans within each document. That removes the need for a separate extraction or compression step before the LLM (LLM Reference).
• Calibrated scores: Cohere says a score of 0.7 should mean a 70% chance the document is useful. It reports expected calibration error of about 0.06, against 0.23–0.38 for competitors. Pro's default threshold stays between 0.47 and 0.53 across domains, which makes fixed relevance cutoffs workable (Cohere).
• Evaluation metric: Cohere scores with its own RCP-nDCG@10 rather than standard nDCG@10, across ViDoRe V3, nanoMTEB, BRIGHT and BEIR (Cohere). These numbers do not compare directly with public leaderboards.

• Ranking: Pro placed first on 39 of 46 datasets and scored 82.2 on BRIGHT, against 76.4 for Voyage AI Rerank 3 (Cohere). AlphaSignal reports this as 12.4 points above Rerank 4 Pro (AlphaSignal).
• The "a third less reading" headline: This is the search-agent test, where Rerank 5 used 35% fewer tokens than a BM25 baseline and accuracy rose from 46% to 58% (Cohere). The 41% figure comes from a separate deep-search benchmark (Cohere).
• Pruning cost: Removing 50% of the context lowered answer quality by 2%; removing 75% lowered it by 7% (Cohere).
• Status of all figures above: Every number is reported by Cohere, and no independent replication was found (AIToolsReview).

Rerank 4 had already taken Cohere to a 32K window in December 2025 (VentureBeat).

The video presents 32K as Rerank 5's jump from the roughly 4K that rerankers offered a year earlier (video 5:45). That is wrong: the context window did not change in this release. What Rerank 5 adds is extraction at passage level and calibrated scores. Voyage AI Rerank 3 is the main competitor named in Cohere's comparisons (Cohere).

• The Context Focus results depend on model judges and on Cohere's own agent harness (AIToolsReview).
• On easy tasks the lead is small, a point or less against competitors on FinQA (AIToolsReview).
• The video does point out that these are the company's own numbers (0:40). It then argues that most business owners haven't heard of Rerank 5 yet and the window to act won't stay open, with no evidence offered (6:34). It also pitches a paid community (2:58).

• Who benefits most: Teams running retrieval-augmented generation (RAG) or agent pipelines with large LLM context bills. Context Focus could cut LLM input tokens with no extra pipeline step. Test it on your own queries first.
• Choosing a tier: Use Fast for agents that issue many searches and Pro for reasoning-heavy, domain-specific documents (Cohere; video 2:21).
• Score thresholds: Calibrated scores could make fixed cutoffs simpler to set. Confirm calibration on in-domain data before relying on them.
• Cost: Prices are low enough that LLM token savings, not reranking fees, will usually decide total cost.