Cohere AI has launched Rerank 5, a new retrieval model available in Pro and Fast tiers that reorders search results before they reach a large language model. The update introduces two specific developer features: Context Focus, which identifies relevant passages within documents, and calibrated relevance scores. For teams building retrieval-augmented generation systems or agents, the goal is straightforward. Send fewer, better tokens to the LLM.
TL;DR
- Size: Parameter count is undisclosed. The model handles a 32K token context window. Two tiers are available:
rerank-v5.0-proandrerank-v5.0-fast. - Runs on: Hosted API, Model Vault, Microsoft Foundry, Amazon SageMaker, private VPC, or on-premises. Speed benchmarks were run on a single NVIDIA A100. There are no open weights or quantized builds.
- Performance: The Pro tier ranks first on 39 of 46 composite datasets in Cohere’s own evaluation.
- Best: Pro scores 82.2 on BRIGHT (RCP-nDCG@10), which is 5.8 points above Voyage AI Rerank 3.
- Worst: Pro median latency is 11.7 seconds per 96-document request, compared to 4.1 seconds for Fast.
What is new in Cohere Rerank 5?
Rerank 5 ships as two distinct models. Pro targets ranking quality on reasoning-heavy queries and long, domain-specific documents. Fast targets low latency and high throughput for agents running many searches.
Both models support over 100 languages and a 32K context window. They also handle semi-structured data such as JSON, tables, and YAML records. Upgrading requires no re-indexing. Developers change the model parameter and keep their existing keyword, vector, or hybrid search setup.
How does Context Focus work?
Context Focus is an optional feature, off by default. A single Rerank call returns document scores plus scored character spans. Each span carries start and end offsets into the original text. Developers set it through the relevant_spans field, with a threshold and max_spans value.
Cohere tested it on 30 QA datasets. Removing half the retrieved text kept 96% of answer quality. Removing three quarters kept 87%. Cohere says teams can push to 10x less context at some quality cost.
In a search-agent test across 17 datasets, BM25 alone scored 46% accuracy. Adding Rerank 5 Pro raised it to 49%. Context Focus at the default 0.1 threshold cut tokens per task by 19% and reached 57%. At 0.4, tokens fell 34% and accuracy hit 58%. Versus BM25 alone, the full stack used 35% fewer tokens. On 6 deep-search benchmarks, tokens fell 41% and accuracy rose from 39% to 52%.
What do calibrated relevance scores change?
Rerank 5 scores read as probabilities. A score of 0.7 means a 70% chance the document is useful.
The Cohere team measured expected calibration error (ECE) against blind human graders and NIST TREC-DL assessors. Rerank 5 scores about 0.06 on both. Voyage AI Rerank 3 scores 0.234 and 0.307. Rerank 4 Pro scores 0.384 and 0.309. In practice, dev teams can filter by threshold instead of a fixed top-k. An agent can also detect when nothing relevant came back.
How does Rerank 5 perform on benchmarks?
All scores below are self-reported by Cohere using RCP-nDCG@10, a Cohere-developed variant of nDCG. Cohere ran the competitor models itself.
Pro leads all four main benchmarks. It scores 92.2 on ViDoRe V3, 90.4 on BEIR, 82.2 on BRIGHT and 87.3 on nanoMTEB. BRIGHT shows the largest jump: 12.4 points over Rerank 4 Pro.
Fast beats Voyage AI Rerank 3 Lite, Qwen3-Reranker-8B and Rerank 4 Pro on all four. It ties Voyage AI Rerank 3 on ViDoRe V3 at 90.6. It trails it on BEIR, 88.1 versus 88.6.
On parsed PDFs, averaged over 47 datasets, Pro scores 89.9 and Fast 88.5. Voyage AI Rerank 3 scores 88.4. Pro ranks first on 14 of 19 non-English languages in Cohere’s internal suite.
How does Rerank 5 compare with competing rerankers?
| Feature | Cohere Rerank 5 Pro | Cohere Rerank 5 Fast | Voyage AI rerank-3 | Voyage AI rerank-3-lite | Qwen3-Reranker-8B |
|---|---|---|---|---|---|
| Parameters | Not disclosed | Not disclosed | Not disclosed | Not disclosed | 8B |
| Context window | 32K | 32K | 32,000 | 32,000 | 32K |
| Languages | 100+ | 100+ | Not disclosed | Not disclosed | 100+ |
| License / access | Proprietary API, VPC, on-prem | Proprietary API, VPC, on-prem | Proprietary API | Proprietary API | Apache 2.0 open weights |
| API price per 1M tokens | $0.05 | $0.02 | $0.05 | $0.02 | Self-hosted |
| Passage extraction | Context Focus | Context Focus | Not disclosed | Not disclosed | Not disclosed |
| BRIGHT* | 82.2 | 77.5 | 76.4 | 69.4 | 69.8 |
| ViDoRe V3* | 92.2 | 90.6 | 90.6 | 87.9 | 85.1 |
| BEIR* | 90.4 | 88.1 | 88.6 | 86.1 | 83.8 |
| nanoMTEB* | 87.3 | 85.1 | 83.9 | 80.7 | 82.1 |
| ECE, human judges* (lower is better) | 0.058 | 0.058 | 0.234 | 0.226 | Not disclosed |
*Scores are RCP-nDCG@10 or ECE from Cohere’s Rerank 5 post, measured by Cohere. Not independently verified.
How can developers deploy Rerank 5?
Rerank 5 is generally available on the Cohere API and Model Vault. It also runs on Microsoft Foundry (Pro, Fast) and Amazon SageMaker (Pro, Fast). It is also available to all North users. API pricing is $0.05 per 1M tokens for Pro and $0.02 for Fast. Model Vault lists $4.00 per instance-hour for both tiers on its Medium tier.
Integrations include LangChain, Haystack, Elasticsearch and Weaviate, per Cohere’s integration docs.
What it means
For teams building agents, the calibrated scores allow filtering by a probability threshold rather than a fixed count. This means an agent can stop searching once it finds enough high-probability results or detect when no relevant data exists. Context Focus reduces the tokens sent to the LLM by 35% to 41% compared to BM25 alone, which lowers costs and speeds up response times.
