Specialist modelActive
Cohere Rerank 5 Pro
Cohere
- Released
- October 21, 2026
- Data date
- October 24, 2026
- Input price
- Per 1M input tokens
Rerank 5 Pro orders text documents by relevance to a search query. Cohere lists about 2 billion parameters and a limit of 32,768 tokens for the query plus each individual document. Context Focus can return relevant text spans.
Specifications and access
| Specification | Value and source |
|---|---|
| Model type | ProprietarySource |
| API model ID | rerank-v5.0-proSource |
| Release | Released on October 21, 2026Source |
| Access | Cohere Rerank API, with additional documented enterprise and cloud deployment optionsSource |
| Deployment options | Cohere API (subject to rate limits), Model Vault, and North; Microsoft Foundry and Amazon SageMaker list Pro and Fast. Private VPC or on-premises deployment is available on request.Source |
| Input | Text, including semi-structured data such as JSON, tables, and records passed as textSource |
| Output | Documents ordered by relevance with a calibrated score from 0 to 1Source |
| Context length | 32,768 tokens per query and individual documentSource |
| Parameter count | Approximately 2 billion parametersSource |
| Languages | Multilingual model; Cohere lists more than 100 languagesSource |
| Context Focus | Can return relevant spans from documents to pass fewer irrelevant tokens to downstream answersSource |
| Deployment options | Cohere API (rate limits apply), Model Vault, Microsoft Foundry and Amazon SageMaker (both list Pro and Fast), North, private VPC or on-premises deployment on requestSource |
| Release | 2026-10-21Source |
| Approximate parameter count in billions | |
| Architecture | proprietarySource |
| Input modality | textSource |
| Source output | ordered documents with calibrated relevance scores from 0 to 1Source |
| Languages | 100+Source |
| Context Focus | optionalSource |
| Price per million tokens in USD | 0.05 tokens (No separate token-side pricing distinction is published in the article.)Source |
| Results across 46 composite evaluations | first on 39; at least second on 43Source |
| Published threshold range | 0.47 to 0.53 for ProSource |
| API model ID evaluated by Exa | Exa's evaluation appendix identifies rerank-v5.0-pro as the Cohere endpoint queried for all seven benchmarks.Source |
| Exa evaluation query date | Protocol says Cohere Rerank 5 Pro was queried on October 21, 2026, for all seven benchmarks.Source |
| Exa evaluation sample sizes | BEIR: 1,646; MIRACL: 12,514; MAIR: 9,972; coding: 317; navigation: 639; freshness: 1,000; authority: 1,000 queries.Source |
| Exa evaluation protocol | Same queries and candidate lists in identical order; BEIR/MIRACL/MAIR use Gemini Embedding 2 top-50 results, coding/navigation use Exa Search Instant top 50, freshness/authority use 60 candidates. (BEIR/MIRACL/MAIR title and document text: 20,000 characters; coding extracts: 720; navigational extracts: 8,000; freshness/authority: 720 plus URL, title and publication date. MAIR uses “Instruct: {instruction}
Query: {query}”.)Source |
| Exa statistical method | nDCG@10; stratified query bootstrap B=10,000, resampling within subsets; macro averages recomputed per draw; paired model differences use the same draws and count as significant only if the paired 95% interval excludes zero.Source |
| Exa evaluation ownership | The article describes its own models as serving Exa production search and names Cohere among competitors.Source |
Benchmark sources at a glance
Each evaluator measures different capabilities under its own conditions. Test series, composite indices and their components stay separate. The detailed results below retain their scales, configurations and source dates.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Coverage is not a quality ranking. Overlapping tests and multiple settings do not count as independent experiments. Vendor claims, historical and disputed results are excluded from this overview and remain labelled in the detailed evidence.
Page 1 of 8
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
Context Focus answer quality: 0% retrieved tokens removed (Show measurement, test conditions, and source)
- Source value
- 55
- Score
- 55
- Metric
- answer judge quality
- Unit
- %
- Category
- retrieval system / context compression
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: Context Focus
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. For each system, rerank the top five from a shared 50-document candidate pool and remove retrieved tokens at the shown Context Focus threshold. Qwen3.5-9B generates answers; Qwen3.5-35B judges them. Mean over 30 single-turn QA datasets from ViDoRe, UNICORN, BRIGHT-Pro, and agentic QA (up to 500 queries per dataset).
Context Focus answer quality: 50% retrieved tokens removed (Show measurement, test conditions, and source)
- Source value
- 53
- Score
- 53
- Metric
- answer judge quality
- Unit
- %
- Category
- retrieval system / context compression
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: Context Focus
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. For each system, rerank the top five from a shared 50-document candidate pool and remove retrieved tokens at the shown Context Focus threshold. Qwen3.5-9B generates answers; Qwen3.5-35B judges them. Mean over 30 single-turn QA datasets from ViDoRe, UNICORN, BRIGHT-Pro, and agentic QA (up to 500 queries per dataset).
Context Focus answer quality: 75% retrieved tokens removed (Show measurement, test conditions, and source)
- Source value
- 48
- Score
- 48
- Metric
- answer judge quality
- Unit
- %
- Category
- retrieval system / context compression
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: Context Focus
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. For each system, rerank the top five from a shared 50-document candidate pool and remove retrieved tokens at the shown Context Focus threshold. Qwen3.5-9B generates answers; Qwen3.5-35B judges them. Mean over 30 single-turn QA datasets from ViDoRe, UNICORN, BRIGHT-Pro, and agentic QA (up to 500 queries per dataset).
Context Focus answer quality: 90% retrieved tokens removed (Show measurement, test conditions, and source)
- Source value
- 41
- Score
- 41
- Metric
- answer judge quality
- Unit
- %
- Category
- retrieval system / context compression
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: Context Focus
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. For each system, rerank the top five from a shared 50-document candidate pool and remove retrieved tokens at the shown Context Focus threshold. Qwen3.5-9B generates answers; Qwen3.5-35B judges them. Mean over 30 single-turn QA datasets from ViDoRe, UNICORN, BRIGHT-Pro, and agentic QA (up to 500 queries per dataset).
Context Focus answer quality: 95% retrieved tokens removed (Show measurement, test conditions, and source)
- Source value
- 35
- Score
- 35
- Metric
- answer judge quality
- Unit
- %
- Category
- retrieval system / context compression
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: Context Focus
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. For each system, rerank the top five from a shared 50-document candidate pool and remove retrieved tokens at the shown Context Focus threshold. Qwen3.5-9B generates answers; Qwen3.5-35B judges them. Mean over 30 single-turn QA datasets from ViDoRe, UNICORN, BRIGHT-Pro, and agentic QA (up to 500 queries per dataset).
Search-agent accuracy: all 17 datasets, without Context Focus (Show measurement, test conditions, and source)
- Source value
- 48.6
- Score
- 48.6
- Metric
- answer accuracy
- Unit
- %
- Category
- end-to-end search agent
- Direction
- Higher is better
- Sample
- 850
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: search-agent stack
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. End-to-end search-agent stack, all 17 datasets, 50 questions per dataset; GLM-5-3-Flash agent, search as the only tool, five results per search, Cohere Rerank 5 Pro reranks the top 100. Configuration: without Context Focus.
- Context
- This is a complete retrieval-and-agent system result, not a standalone reranker benchmark.
Search-agent accuracy: all 17 datasets, Context Focus threshold 0.1 (Show measurement, test conditions, and source)
- Source value
- 56.6
- Score
- 56.6
- Metric
- answer accuracy
- Unit
- %
- Category
- end-to-end search agent
- Direction
- Higher is better
- Sample
- 850
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: search-agent stack
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. End-to-end search-agent stack, all 17 datasets, 50 questions per dataset; GLM-5-3-Flash agent, search as the only tool, five results per search, Cohere Rerank 5 Pro reranks the top 100. Configuration: Context Focus threshold 0.1.
- Context
- This is a complete retrieval-and-agent system result, not a standalone reranker benchmark.
Search-agent accuracy: all 17 datasets, Context Focus threshold 0.4 (Show measurement, test conditions, and source)
- Source value
- 58.3
- Score
- 58.3
- Metric
- answer accuracy
- Unit
- %
- Category
- end-to-end search agent
- Direction
- Higher is better
- Sample
- 850
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: search-agent stack
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. End-to-end search-agent stack, all 17 datasets, 50 questions per dataset; GLM-5-3-Flash agent, search as the only tool, five results per search, Cohere Rerank 5 Pro reranks the top 100. Configuration: Context Focus threshold 0.4.
- Context
- This is a complete retrieval-and-agent system result, not a standalone reranker benchmark.
Search-agent accuracy: six deep-search datasets, without Context Focus (Show measurement, test conditions, and source)
- Source value
- 44
- Score
- 44
- Metric
- answer accuracy
- Unit
- %
- Category
- end-to-end search agent
- Direction
- Higher is better
- Sample
- 300
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: search-agent stack
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. End-to-end search-agent stack, six deep-search datasets, 50 questions per dataset; GLM-5-3-Flash agent, search as the only tool, five results per search, Cohere Rerank 5 Pro reranks the top 100. Configuration: without Context Focus.
- Context
- This is a complete retrieval-and-agent system result, not a standalone reranker benchmark.
Search-agent accuracy: six deep-search datasets, Context Focus threshold 0.1 (Show measurement, test conditions, and source)
- Source value
- 49.7
- Score
- 49.7
- Metric
- answer accuracy
- Unit
- %
- Category
- end-to-end search agent
- Direction
- Higher is better
- Sample
- 300
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: search-agent stack
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. End-to-end search-agent stack, six deep-search datasets, 50 questions per dataset; GLM-5-3-Flash agent, search as the only tool, five results per search, Cohere Rerank 5 Pro reranks the top 100. Configuration: Context Focus threshold 0.1.
- Context
- This is a complete retrieval-and-agent system result, not a standalone reranker benchmark.
Search-agent accuracy: six deep-search datasets, Context Focus threshold 0.4 (Show measurement, test conditions, and source)
- Source value
- 52
- Score
- 52
- Metric
- answer accuracy
- Unit
- %
- Category
- end-to-end search agent
- Direction
- Higher is better
- Sample
- 300
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: search-agent stack
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. End-to-end search-agent stack, six deep-search datasets, 50 questions per dataset; GLM-5-3-Flash agent, search as the only tool, five results per search, Cohere Rerank 5 Pro reranks the top 100. Configuration: Context Focus threshold 0.4.
- Context
- This is a complete retrieval-and-agent system result, not a standalone reranker benchmark.
RCP-nDCG@10: ViDoRe V3 (Show measurement, test conditions, and source)
- Source value
- 92.2
- Score
- 92.2
- Metric
- RCP-nDCG@10 retrieval quality
- Unit
- points
- Category
- headline
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Cohere benchmark chart: headline benchmarks
- Status
- active
- Retrieved at
- 2026-10-24
- Methodology
- Vendor-reported result. RCP-nDCG@10, mean over public benchmark datasets; ViDoRe V3 has 8 datasets, BEIR 13, BRIGHT 12, and nanoMTEB 13.
Page 1 of 8
Published prices
- API price
- USD 0.05 per 1 million tokensSource
Sources and data date
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 24, 2026. 111 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Unresolved | Linked full benchmark spreadsheet Note The official article links a spreadsheet with the full benchmark results, but the live sheet returned HTTP 401. The readable article tables and linked public charts are included; unseen spreadsheet rows or metadata cannot be claimed complete. |
| Unresolved | Full multilingual benchmark results Note The article says its internal multilingual suite covers 20 languages and 74 datasets but publicly plots six language aggregates. The linked full-results sheet returned HTTP 401. |
| Unresolved | Numeric confidence interval bounds Note Exa describes 95% confidence intervals and shows error bars, but the readable article does not give numerical endpoints per Cohere score. The linked sheet and its ordinary public CSV exports returned HTTP 401. |
| Not found in the inspected sources | MTEB reranker scores for exact Cohere Rerank 5 identity Note MTEB model lists and exact lookup endpoints did not expose rerank-v5.0-pro or Cohere/rerank-v5.0-pro; no score is assigned. |
| Not found in the inspected sources | RTEB score rows for exact Cohere Rerank 5 identity Note All nine published RTEB score tables were searched. Cohere embedding rows are not Rerank 5 results; neither exact target alias was found. |
| Additional source | Cohere Rerank 5 announcement (retrieved October 24, 2026) · Cohere Rerank 5 announcement · Editorial description reviewed October 24, 2026 |
| Additional source | Cohere Rerank API documentation (retrieved October 24, 2026) · Cohere Rerank documentation · Editorial description reviewed October 24, 2026 |
| Additional source | Cohere Rerank 5 changelog (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: headline (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: finance (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: healthcare (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: scientific (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: energy (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: parsed PDFs (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: throughput and latency (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: Context Focus (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: search agent (retrieved October 24, 2026) |
| Additional source | Cohere benchmark chart: score calibration (retrieved October 24, 2026) |
| Additional source | Exa SOTA reranking evaluation (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |
| Additional source | Model metadata (retrieved October 24, 2026) |