Cohere releases Rerank 5 with Context Focus passage selection

• Rerank 5 is available in Pro and Fast tiers, according to Cohere's announcement.
• Context Focus selects scored passages within documents; the 41% token-saving claim in Cohere's launch thread comes from six deep-search benchmarks.
• Fast delivers nearly three times Pro's throughput and faster responses in Cohere's tests.

Context Focus's Unicode character offsets are a welcome bit of API plumbing: applications can slice passages directly from source documents. Cohere's model docs disclose roughly 2B parameters for Pro and 500M for Fast, both with 32,768-token context windows and text-only input.

Passage selection happens inside the existing POST /v2/rerank call. The API reference documents this optional request object:

• Enablement: Context Focus is off by default.
• Threshold: The default is 0.1; higher values filter out more text.
• Span limit: max_spans caps passages per document. Omitting it returns all qualifying spans.
• Response: Each span contains start, exclusive end, and score. A result with no qualifying passage omits relevant_spans.

Broad summarization still needs full-document context, according to the usage guide, which also flags less clean passage boundaries in tables, forms and code.

Cohere's search-agent evaluation used glm-5-3-flash, five results per search and 50 questions per dataset. The reported comparisons started with BM25 keyword search and ended with Rerank 5 Pro plus Context Focus at a 0.4 threshold:

• All 17 datasets: Accuracy rose from 46% to 58%, with 35% fewer tokens per task.
• Six deep-search benchmarks: Accuracy rose from 39% to 52%, with 41% fewer tokens.

The token metric totals every LLM call in an agent run and includes reasoning and cached tokens. In a separate 30-dataset QA experiment, removing half the retrieved text cost two percentage points of answer quality.

Pro ranked first on 39 of 46 component datasets and at least second on 43 in Cohere's evaluation. Its headline results were:

• BRIGHT: 82.2, beating Rerank 4 Pro by 12.4 points and Voyage AI Rerank 3 by 5.8.
• ViDoRe V3: 1.6 points ahead of the strongest external competitor tested.
• BEIR: 1.8 points ahead.
• nanoMTEB: 3.4 points ahead.

These are vendor-run results using RCP-nDCG@10, a Cohere-developed metric intended to address incomplete relevance judgments. The PDF retrieval evaluation used extracted text, with documents parsed by Gemini 1.5 Flash.

Both tiers keep the existing relevance_score response field. Cohere interprets those scores as probabilities that a document is useful in its calibration evaluation:

• Useful more often than not: The observed cutoff was 0.50 for Pro and 0.45 for Fast.
• Blind human judgments: Pro's expected calibration error was 0.058, versus 0.384 for Rerank 4 Pro.
• NIST TREC-DL judgments: Pro's error was 0.057, versus 0.309 for Rerank 4 Pro.

Those evaluations covered 2,262 blind-graded documents and 20,646 documents assessed by NIST. A relevance cutoff can leave an agent with no qualifying documents, providing a trigger for another search instead of always supplying a fixed top-k.

API pricing is $0.05 per million tokens for Pro and $0.02 for Fast, according to the launch post. Both sit above existing keyword, vector or hybrid retrieval without re-indexing.

Private VPC and on-premises deployments are available by arrangement. Cohere's hosted availability announcement covers both tiers:

Fast's relative speedup comes from a specific load test:

• Hardware: One NVIDIA A100 GPU.
• Requests: 96 documents per request, with concurrency of eight.
• Sample: 1,000 requests per document length.
• Throughput: About 187 documents per second at roughly 1,000 tokens each, 2.9 times Pro's throughput.