Last updated: October 31, 2026
Cohere Rerank 5, released October 21, 2026, comes in Pro and Fast tiers for reordering search results before they reach an LLM. It adds Context Focus, which returns the most relevant passages inside each document, and calibrated scores you can threshold. API pricing is $0.05 (Pro) or $0.02 (Fast) per million tokens.
This guide draws on Cohere’s launch post, the Rerank 5 changelog entry, the Rerank model documentation, the Rerank API reference and the Cohere pricing page. For background, see our answer on the Cohere and Aleph Alpha merger agreement.
Contents
- What is new in Cohere Rerank 5
- Specs and pricing at a glance
- How to call Cohere Rerank 5
- Using Context Focus
- Calibrated scores and thresholds
- Benchmarks
- Rerank 5 vs Rerank 4
- Limitations
- Frequently asked questions
What is new in Cohere Rerank 5
A reranker reorders candidates from your first-stage search by relevance. According to the changelog, Rerank 5 brings four changes that matter to builders:
- Two tiers.
rerank-v5.0-protargets the highest quality;rerank-v5.0-fasttargets low latency and high throughput. - Context Focus. An optional setting that returns the most relevant passages in each document as scored character spans, in the same call.
- Calibrated relevance scores. Scores stay consistent across queries and domains, so one threshold can decide what reaches your LLM.
- Broad input support. Over 100 languages, a 32k-token context, and semi-structured records such as JSON or tables passed as YAML strings.
Specs and pricing at a glance
| Item | Rerank 5 Pro | Rerank 5 Fast |
|---|---|---|
| Release date | October 21, 2026 | October 21, 2026 |
| API model name | rerank-v5.0-pro |
rerank-v5.0-fast |
| Parameters | About 2B | About 500M |
| Context length | 32,768 tokens (query plus document) | 32,768 tokens (query plus document) |
| Input | Text; structured data as YAML strings | Text; structured data as YAML strings |
| API price per 1M tokens | $0.05 | $0.02 |
| Model Vault (Medium) | $4.00 per hour or $2,500 per month | $4.00 per hour or $2,500 per month |
| Best for | Reasoning-heavy queries, long or domain-specific documents | High-volume search and agent loops |
Specs are from the model documentation, token prices from the launch post and Model Vault rates from the pricing page. Both tiers are available on the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker and to Cohere North users. Microsoft’s agent tooling on Foundry is covered in our Microsoft Copilot harness guide.
How to call Cohere Rerank 5
Rerank uses one endpoint, POST https://api.cohere.com/v2/rerank. The steps below follow the API reference.
- Create an API key in your Cohere account and export it as
CO_API_KEY. - Run your existing keyword, vector or hybrid search and collect the candidate documents as a list of strings.
- Send
model,queryanddocumentsin the request body, with aBearertoken in theAuthorizationheader. Cohere recommends no more than 1,000 documents per request. - Optionally set
top_nto limit how many results come back, andmax_tokens_per_doc(default 4,096) to control truncation of long documents. - Read the
resultslist. Each item has anindexpointing back to your original list and arelevance_scorebetween 0 and 1. - Pass only the documents that clear your score threshold to the LLM.
A newer optional field, priority, controls how early a request is handled under load. Lower numbers mean higher priority, and 0, the default, is the highest. To upgrade from Rerank 4 or 3.5, change only the model value; Cohere says no re-indexing is needed.
Using Context Focus
Context Focus is the headline feature. It lets you send your LLM only the passages that answer the query. Turn it on by adding this object to a Rerank 5 request: "relevant_spans": {"mode": "enabled", "threshold": 0.3, "max_spans": 10}.
thresholddefaults to 0.1. A higher value removes more text.max_spanscaps spans per document; if unset, all qualifying spans return.- Each span has
startandendoffsets, counted in Unicode codepoints, plus ascore. - If no passage meets the threshold,
relevant_spansis left out of that result.
Cohere’s launch post reports that removing half of the retrieved text kept 96% of answer quality, and removing three quarters kept 87%. In a search-agent test, accuracy rose from 46% with BM25 alone to 49% with Rerank 5 Pro. Context Focus at the default 0.1 threshold cut tokens per task by 19% and lifted accuracy to 57%; at 0.4, tokens fell 34% and accuracy reached 58%.
Calibrated scores and thresholds
Cohere says a Rerank 5 score of 0.7 now means about a 70% chance the document is useful. The “more likely useful than not” point is 0.50 for Pro and 0.45 for Fast. Cohere reports an expected calibration error of about 0.058 for both tiers, against 0.384 for Rerank 4 Pro on human judgments.
In practice, filter by score instead of a fixed top-k, and when nothing clears the threshold, have an agent rewrite the query or say it does not know. Cohere found the Pro default stays between 0.47 and 0.53 across its test sets, so one threshold can serve several domains.
Benchmarks
These are Cohere’s own figures from the launch post, measured with its RCP-nDCG@10 method. Rerank 5 Pro ranked first on 39 of 46 composite datasets. On BRIGHT, a reasoning-heavy retrieval benchmark, it scored 82.2, which is 5.8 points above Voyage AI Rerank 3 and 12.4 points above Rerank 4 Pro. On parsed PDFs, Pro averaged 89.9 and Fast 88.5, both ahead of Rerank 4 Pro at 88.3. Rerank 5 Fast reranks about 187 documents of 1,000 tokens per second on one NVIDIA A100, with 2.9 times the throughput of Pro.
Rerank 5 vs Rerank 4
| Feature | Rerank 4 | Rerank 5 |
|---|---|---|
| Model names | rerank-v4.0-pro, rerank-v4.0-fast |
rerank-v5.0-pro, rerank-v5.0-fast |
| Context length | 32k | 32k |
| Context Focus (relevant spans) | No | Yes |
| Calibrated scores | No | Yes |
| Calibration error (human judgments, Pro) | 0.384 | 0.058 |
| BRIGHT score (Pro) | 12.4 points lower | 82.2 |
The model documentation lists Context Focus and calibrated scores only for the v5 models, and the API reference says relevant_spans works only with Rerank v5 and newer. If you build retrieval for agents, compare this with our Perplexity Decisions API guide, and for a low-cost model to read the reranked context, see our Claude Haiku 5.5 guide or MiniMax M3.1 Flash guide.
Limitations
- Rerank 5 accepts text only.
- It does not support instructions that change how relevance is judged.
- Context Focus works best for targeted queries. Summarization needs full documents, and tables, forms and code split less cleanly than prose.
- Context Focus is available through the Cohere API; the docs do not confirm it on other platforms.
- All benchmark figures come from Cohere.
Frequently asked questions
What is Cohere Rerank 5?
Cohere Rerank 5 is a family of reranking models released on October 21, 2026, in two tiers: rerank-v5.0-pro and rerank-v5.0-fast. They reorder search results by relevance to a query and add Context Focus, which returns the most relevant passages inside each document, plus calibrated relevance scores.
How much does Cohere Rerank 5 cost?
Cohere lists API pricing of $0.05 per million tokens for Rerank 5 Pro and $0.02 per million tokens for Rerank 5 Fast. For dedicated capacity, Cohere’s pricing page lists Model Vault instances for either model at $4.00 per hour or $2,500 per month on the Medium tier.
How do I turn on Context Focus?
Add a relevant_spans object to a Rerank 5 request with mode set to enabled. You can also set threshold, which defaults to 0.1, and max_spans. Each result then includes scored passages with start and end character offsets. Context Focus is off by default.
Do I need to re-index my documents to upgrade to Rerank 5?
No. Rerank reads raw text at query time, so upgrading only requires changing the model parameter to rerank-v5.0-pro or rerank-v5.0-fast. Your existing keyword, vector or hybrid search index stays the same.
What does a Rerank 5 relevance score mean?
Cohere says Rerank 5 scores are calibrated, so a score of 0.7 means about a 70% chance the document is useful. The point where a document is more likely useful than not is 0.50 for Pro and 0.45 for Fast.
Should I use Rerank 5 Pro or Rerank 5 Fast?
Use Pro when ranking quality matters most, such as complex, reasoning-heavy queries and long or domain-specific documents. Use Fast for high-volume search and agent loops. Cohere reports Fast has about 2.9 times the throughput of Pro and nearly three times faster response times.
Sources
- Cohere: Introducing Rerank 5
- Cohere Docs: Announcing Cohere’s Rerank 5 Models
- Cohere Docs: Cohere Rerank
- Cohere Docs: Rerank API (v2)
- Cohere: Pricing
Image: Cohere logo, CC0, by Ayub MK09 via Wikimedia Commons. Placed on a white background and resized. Cohere is a trademark of its owner. FreeAnswers is not affiliated with or endorsed by Cohere.
1 thought on “Cohere Rerank 5: Power-User Guide to Context Focus, Pricing and the API”