What it is
Rerank 5 is Cohere's current reranker, released on 21 October 2026 as rerank-v5.0-pro and rerank-v5.0-fast. A reranker rescores the shortlist a search system returns by reading the query and each document together, which is slower than vector search but more precise. Cohere puts Pro at about 2 billion parameters and Fast at about 500 million; the weights are not released.
Both accept 32,768 tokens per query-and-document pair, text only, in more than 100 languages. Two features are new. Scores are calibrated, so 0.7 is meant to mean a 70% chance the document is useful, and an optional Context Focus mode returns the scored character spans inside each document that answer the query, so only those need reach the LLM.
Unlike Rerank 4, which is billed per search, Rerank 5 is billed per token: $0.05 per million for Pro and $0.02 for Fast. It is on Cohere's API, Model Vault, Microsoft Foundry, Amazon SageMaker and private deployments.
Our review
Calibration is the change that matters most, more than the benchmark lead. Earlier rerankers returned scores that only meant something relative to each other, so teams hand-tuned cut-offs per corpus. Cohere reports an expected calibration error of 0.058 for both Rerank 5 sizes against 0.384 for Rerank 4 Pro and 0.234 for Voyage's rerank-3, and says Pro's sensible threshold stays between 0.47 and 0.53 across BRIGHT, NanoBEIR and ViDoRe V3. If that holds on your data, a single number can decide when an assistant says it does not know.
On quality the launch post claims a clear lead: first on 39 of 46 composite datasets, 82.2 on the reasoning-heavy BRIGHT benchmark, 5.8 points ahead of Voyage rerank-3 and 12.4 ahead of Rerank 4 Pro. All of it is measured by Cohere with its own RCP-nDCG@10 metric, and we found no independent results two days after launch. The margins on parsed PDFs are thin, 89.9 against Voyage's 88.4, so the gain will depend on the corpus.
The price undercuts its predecessor for typical retrieval. A call with 20 passages of 300 tokens is about 6,000 tokens, roughly $0.0003 on Pro, against a flat $0.0025 per search for Rerank 4 Pro on Vercel's gateway. Long documents erase the gap: 100 documents of 500 tokens costs the same either way. Fast, at $0.02 per million, handles about 187 documents of 1,000 tokens a second on one A100, 2.9 times Pro's rate, by Cohere's account.
Context Focus is the feature to test before relying on it. Cohere says cutting half the context kept 96% of answer quality, but its own docs warn that tables, forms and code split less cleanly than prose, and the mode is available only on Cohere's API, not on Foundry or SageMaker. The docs also say the model does not follow instructions that change how relevance is judged, so domain-specific notions of relevance still have to come from the query itself.
Where it earns its place
A search engineer on a support assistant answering 2 million queries a month
Cut wrong answers by refusing when nothing relevant is retrieved
Rerank the top 20 hybrid-search results with Pro and fall back to 'I don't know' below a score near 0.5, Cohere's reported break-even point. At 300-token passages that is about 12 billion tokens, roughly $600 a month on Pro or $240 on Fast.
An AI engineer whose RAG prompts are too long and too costly
Send the LLM only the parts of each document that answer the question
Enable relevant_spans with a threshold of about 0.3 and pass only the returned spans to the generator. Cohere reports 50% less context at 96% of answer quality; tables and code need checking, and the mode works only on Cohere's own API.
An IT lead at a bank that cannot send documents to a multi-tenant API
Add reranking inside the bank's own cloud account
Deploy through Amazon SageMaker, Microsoft Foundry or a private VPC deployment, where Cohere says it receives no inputs or outputs. A dedicated Model Vault instance is listed at $4 an hour, about $2,500 a month, which suits steady volume better than bursty traffic.
Before you commit
- Setup
- Same Rerank endpoint as earlier versions: pass rerank-v5.0-pro or rerank-v5.0-fast with a query and documents; no re-indexing is needed. Pipelines moving from Rerank 4 must re-fit score thresholds, since the scale has changed.
- Integrations
- Cohere API and SDKs, Model Vault, Microsoft Foundry, Amazon SageMaker, Cohere North and private VPC or on-premises deployment; framework support in LangChain, Haystack, Elasticsearch and Weaviate.
- Data and privacy
- On Cohere's SaaS, prompts and outputs are logged and deleted after 30 days; training use can be switched off under Data Controls, and zero data retention is available to approved enterprise customers. On cloud platforms and private deployments Cohere receives no inputs. SOC 2 Type II and ISO 27001.
- Cost at scale
- Pro at $0.05 and Fast at $0.02 per million tokens: 12 billion tokens a month is about $600 or $240. A Model Vault Medium instance is $4.00 an hour, $2,500 a month, for either size.
- Limits
- 32,768 tokens per query-document pair, with longer documents chunked and scored in several passes. Production keys allow 1,000 rerank requests a minute, trial keys 10. Text only; no relevance-steering instructions.
What's good
- Calibrated scores make one threshold usable across corpora
- Per-token pricing: $0.05 per million on Pro, $0.02 on Fast
- Context Focus returns the relevant spans in each document
- 32K tokens per pair; Cohere API, Foundry, SageMaker or private deployment
What's not
- Every published benchmark is Cohere's own
- Context Focus only on Cohere's API
- Text only; images and PDFs must be parsed first
- Ignores instructions that redefine relevance
Best for
Teams running production RAG or enterprise search who want a hosted reranker with dependable score thresholds and private-deployment options.
Alternatives
Cohere Rerank 5: common questions
- Is Cohere Rerank 5 worth it?
- Rerank 5 makes reranking cheap enough to stop rationing and calibrated enough to trust a threshold, though every benchmark behind it so far is Cohere's own. We rate it 4.4 out of 5.
- How much does Cohere Rerank 5 cost?
- Pro $0.05 and Fast $0.02 per million tokens on Cohere's API. Model Vault dedicated instance $4.00/hour ($2,500/month) for either size.
- What are the downsides of Cohere Rerank 5?
- Every published benchmark is Cohere's own; context Focus only on Cohere's API; text only; images and PDFs must be parsed first; ignores instructions that redefine relevance.
- Who is Cohere Rerank 5 best for?
- Teams running production RAG or enterprise search who want a hosted reranker with dependable score thresholds and private-deployment options.
- How much does Cohere Rerank 5 cost for a team or in production?
- Pro at $0.05 and Fast at $0.02 per million tokens: 12 billion tokens a month is about $600 or $240. A Model Vault Medium instance is $4.00 an hour, $2,500 a month, for either size.
- Is Cohere Rerank 5 safe for company data?
- On Cohere's SaaS, prompts and outputs are logged and deleted after 30 days; training use can be switched off under Data Controls, and zero data retention is available to approved enterprise customers. On cloud platforms and private deployments Cohere receives no inputs. SOC 2 Type II and ISO 27001.
- What are the best Cohere Rerank 5 alternatives?
- The closest alternatives we have reviewed are Voyage rerank-3, Cohere Rerank 4, zerank-2, Jina Reranker v3.5 and mxbai-rerank v2.
Answers checked Oct 23, 2026.
Make Cohere Rerank 5? Show this rating on your site
