AI Tools & Business / AI news for Malaysia
Cohere Releases Rerank 5 With Context Focus for Search and Agents
Cohere has released two reranking models that score search results and, optionally, cut each document down to relevant passages. The claimed savings are large, but they come from Cohere’s own tests.
In brief
- Cohere released Rerank 5 on 21 October in two tiers, Pro for ranking quality and Fast for lower latency and cost.[1][2]
- An optional Context Focus feature scores passages inside each document so an application can send an AI model less text.[1]
- Cohere lists API prices of US$0.05 per million tokens for Pro and US$0.02 for Fast; its benchmark results are its own and Malay is not named.[1]
What Cohere released
Cohere has released Rerank 5, a model that re-orders search results by relevance before an AI system reads them. It comes in two tiers, Pro and Fast, and adds an optional feature called Context Focus. Cohere announced it on 21 October.[1][2]
Most of the evidence is Cohere’s own testing, and the post names no customers.[1]
What a reranker does in AI search
Many company assistants retrieve candidate documents, then ask a model to answer from them. A reranker sits between the two steps and scores each candidate against the question. Cohere says Rerank reads raw text at query time, so teams do not need to re-index or re-embed their data. Upgrading means changing the model parameter in the request.[1][2]
Rerank 5 Pro is aimed at reasoning-heavy queries and long or specialised documents. Rerank 5 Fast targets low latency and high throughput. Cohere reports about 187 documents a second for Fast on one NVIDIA A100 GPU, 2.9 times Pro’s rate.[1][2]
Context Focus trims what the model reads
Context Focus is optional and off by default. It scores passages within each document, so an application can pass only the relevant text to the model. In Cohere’s chart, removing half the retrieved text kept 96% of answer quality, and removing three-quarters kept 87%.[1][2]
In a 17-dataset search agent test, Pro lifted accuracy from 46% to 49% over a BM25 keyword baseline. With Context Focus set to 0.4, accuracy reached 58% and token use fell 35% against BM25 alone. Cohere also says its scores are calibrated, so each reads as the estimated chance that a document is useful.[1]
Benchmark claims come from Cohere
Cohere says Pro ranked first on 39 of 46 composite datasets. On BRIGHT, a reasoning-heavy retrieval test, Pro scored 82.2, which Cohere says is 5.8 points above Voyage AI Rerank 3 and 12.4 above Rerank 4 Pro. The post uses Cohere’s own metric, RCP-nDCG@10, and does not mention independent verification.[1]
For languages, Cohere reports that Pro ranked first on 14 of 19 non-English languages in an internal suite. Its table names six languages, and Malay and other Southeast Asian languages are not among them.[1]
Prices and where it runs
Cohere lists Pro at US$0.05 per million tokens and Fast at US$0.02. Both are available through the Cohere API and Model Vault, its single-tenant option. Cohere also lists Microsoft Foundry, Amazon SageMaker and North, and says private deployments go through sales.[1][2]
The documentation lists a 32,000-token context length and support for more than 100 languages. It gives no rates for private deployments.[1][2]
Why Malaysia should care
A Malaysian company building a chat assistant over its policies or price lists could use a reranker to pick the passages that reach the model. The announcement does not name Malay among its tested languages, and Cohere’s multilingual results rest on an internal suite of 19 non-English languages. Cohere’s documentation says the models cover more than 100 languages, which is not a published Malay result. Prices are in US dollars, and cloud listings do not confirm a Malaysian region.
Teams building document assistants
A reranker can improve which passages reach the model and may reduce the text sent.[1][2]
Practical move: Compare Rerank 5 with your current ranking on questions your staff already answered.
Companies working in Malay and mixed languages
Cohere does not name Malay in its published results.[1]
Practical move: Build a small Malay and English test set before moving production traffic.
What Malaysians can do now
- List 30 real staff questions with the documents that answer them, including Malay and mixed-language examples.
- Test Rerank 5 Fast and Pro against your current ranking, then compare answer quality, delay and cost.
- Try Context Focus at more than one threshold, and check that trimmed text keeps legal or price details.
What we still do not know
Cohere has not answered several questions Malaysian buyers may have.
- How Rerank 5 performs on Malay, Malay-English mixed text and Chinese or Tamil business documents.
- Whether the results hold in independent tests, since the announcement relies on Cohere’s own metric.
- Which cloud regions serve Malaysian accounts and what private deployment costs.
Sources
- 1.Cohere Rerank 5 | More Relevant Context for RAG and Agents Cohere, 21 October 2026
- 2.Rerank 5 Models Cohere Documentation
- 3.File:Cern datacenter.jpg Wikimedia Commons, 12 March 2010
- 4.File:Book shelves of Bibliothek Medizin, Universitätsbibliothek Bern.jpg Wikimedia Commons, 23 March 2022
- 5.File:Bukit Bintang Crossing in 2022.jpg Wikimedia Commons, 30 January 2022
More AI News
View all
