Cohere Releases Rerank 5 Retrieval Models

Cohere released Rerank 5, a pair of enterprise search reranking models called Pro and Fast, through its API and several cloud platforms. Pro leads the public benchmarks Cohere tested, while Fast offers near-leading performance at lower cost and latency. Both models add Context Focus, an optional feature that returns the most relevant text spans within documents alongside document relevance scores.

On 46 composite datasets, Pro ranked first on 39 and placed at least second on 43. Its RCP-nDCG@10 score on BRIGHT, a benchmark for reasoning-intensive retrieval, was 82.2—5.8 points above Voyage AI Rerank 3 and 12.4 above Rerank 4 Pro. Pro also led the strongest external competitor by 1.6 points on ViDoRe V3, 1.8 on BEIR and 3.4 on nanoMTEB. Fast outperformed Voyage AI Rerank 3 Lite and both Qwen3-Reranker-0.6B and Qwen3-Reranker-8B across all four headline benchmarks.

In the six selected business languages, Pro scored 91.2 in Spanish, 90.5 in French, 90.6 in German, 89.7 in Arabic, 89.4 in Japanese and 91.8 in Korean. It improved on Rerank 4 Pro by at least two points in Spanish, French, German, Arabic and Korean, and ranked first in 14 of 19 non-English languages in Cohere’s internal suite. On parsed PDF retrieval, Pro averaged 89.9 across 47 datasets, ahead of Voyage AI Rerank 3 at 88.4 and Rerank 4 Pro at 88.3; Fast scored 88.5. Under sustained load on one NVIDIA A100, Fast reranked about 187 standard-length documents of 1,000 tokens per second, with 2.9x Pro’s throughput and nearly 3x faster response times.

Cohere says Context Focus can remove half of retrieved text with a 2 percentage point reduction in answer quality, retaining 96%; removing three-quarters reduced quality by 7 points, retaining 87%. At those settings, teams can trade quality for up to 10x less context. In a search-agent evaluation across 17 datasets, adding Pro to BM25 keyword search raised accuracy from 46% to 49%. Context Focus at the default 0.1 threshold reduced tokens per task by 19% and raised accuracy to 57%; at 0.4, tokens fell 34% and accuracy reached 58%. The full setup at 0.4 used 35% fewer tokens than BM25 alone, with accuracy of 58% versus 46%. Across six deep-search benchmarks, it used 41% fewer tokens as accuracy rose from 39% to 52%.

Rerank 5’s relevance scores are calibrated to approximate the probability that a document is useful: a score of 0.7 represents a 70% chance. Pro and Fast use respective useful-document thresholds of 0.50 and 0.45. Against two independent human-judgment sets, Pro and Fast each recorded expected calibration error of 0.058 on the human set; on NIST judgments, Pro scored 0.057 and Fast 0.058. Cohere lists API prices of $0.05 per 1 million tokens for Pro and $0.02 for Fast. Both are generally available through Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker and North; private VPC and on-premises deployments are also offered. The models work with keyword, vector or hybrid search without re-indexing.

Cohere released Rerank 5, a pair of enterprise search reranking models called Pro and Fast, through its API and several cloud platforms. Pro leads the public benchmarks Cohere tested, while Fast offers near-leading performance at lower cost and latency. Both models add Context Focus, an optional feature that returns the most relevant text spans within documents alongside document relevance scores.

On 46 composite datasets, Pro ranked first on 39 and placed at least second on 43. Its RCP-nDCG@10 score on BRIGHT, a benchmark for reasoning-intensive retrieval, was 82.2—5.8 points above Voyage AI Rerank 3 and 12.4 above Rerank 4 Pro. Pro also led the strongest external competitor by 1.6 points on ViDoRe V3, 1.8 on BEIR and 3.4 on nanoMTEB. Fast outperformed Voyage AI Rerank 3 Lite and both Qwen3-Reranker-0.6B and Qwen3-Reranker-8B across all four headline benchmarks.

In the six selected business languages, Pro scored 91.2 in Spanish, 90.5 in French, 90.6 in German, 89.7 in Arabic, 89.4 in Japanese and 91.8 in Korean. It improved on Rerank 4 Pro by at least two points in Spanish, French, German, Arabic and Korean, and ranked first in 14 of 19 non-English languages in Cohere’s internal suite. On parsed PDF retrieval, Pro averaged 89.9 across 47 datasets, ahead of Voyage AI Rerank 3 at 88.4 and Rerank 4 Pro at 88.3; Fast scored 88.5. Under sustained load on one NVIDIA A100, Fast reranked about 187 standard-length documents of 1,000 tokens per second, with 2.9x Pro’s throughput and nearly 3x faster response times.

Cohere says Context Focus can remove half of retrieved text with a 2 percentage point reduction in answer quality, retaining 96%; removing three-quarters reduced quality by 7 points, retaining 87%. At those settings, teams can trade quality for up to 10x less context. In a search-agent evaluation across 17 datasets, adding Pro to BM25 keyword search raised accuracy from 46% to 49%. Context Focus at the default 0.1 threshold reduced tokens per task by 19% and raised accuracy to 57%; at 0.4, tokens fell 34% and accuracy reached 58%. The full setup at 0.4 used 35% fewer tokens than BM25 alone, with accuracy of 58% versus 46%. Across six deep-search benchmarks, it used 41% fewer tokens as accuracy rose from 39% to 52%.

Rerank 5’s relevance scores are calibrated to approximate the probability that a document is useful: a score of 0.7 represents a 70% chance. Pro and Fast use respective useful-document thresholds of 0.50 and 0.45. Against two independent human-judgment sets, Pro and Fast each recorded expected calibration error of 0.058 on the human set; on NIST judgments, Pro scored 0.057 and Fast 0.058. Cohere lists API prices of $0.05 per 1 million tokens for Pro and $0.02 for Fast. Both are generally available through Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker and North; private VPC and on-premises deployments are also offered. The models work with keyword, vector or hybrid search without re-indexing.