Cohere Launches Embed 5 and New Search Metric

Pro and Fast models share a vector space to lower query costs, while the company challenges traditional retrieval benchmarks with in-house math.

On September 30, 2026, Cohere commercially launched its Embed 5 model family, targeting enterprise search and retrieval-augmented generation (RAG) pipelines, where models look up external documents before answering queries. The release includes Embed 5 Pro and Embed 5 Fast, paired with an unusual technical pitch: a new mathematical evaluation framework named RCP-nDCG@10. Cohere claims standard industry benchmarks fail to measure modern retrieval accurately. Both models are available on Cohere’s API, Model Vault, Microsoft Foundry, and Amazon SageMaker, priced at $0.12 per million tokens for Pro and $0.08 for Fast.

Cutting query costs with Embed 5

The main engineering shift centers on a concept Cohere calls “Index Once, Query Cheap.” In semantic search systems, turning documents into embeddings (numerical coordinates that let computers match text and images by meaning) takes serious compute time and budget. When engineering teams want to switch to a lighter model to lower runtime query latency and infrastructure bills, they usually have to re-index their entire document archive from scratch.

Embed 5 addresses that bottleneck by placing Pro and Fast in the exact same vector space. Teams can index their primary document archives once with Embed 5 Pro, then use Embed 5 Fast exclusively to process incoming user queries in production without rebuilding the index. Both models support context windows up to 128,000 tokens, handle more than 100 languages, and take multimodal inputs such as scanned documents and diagrams. Cohere kept support for Matryoshka representations, letting developers compress 2,048-dimensional vectors down to 256 dimensions in float, int8, or binary formats to save storage space without stripping out core semantic data.

A calibrated rubric to replace static labels

Alongside the model launch, Cohere sparked a debate on how the industry measures retrieval accuracy. Established evaluation suites like the Massive Text Embedding Benchmark (MTEB) rely on static document lists with rigid human labels. That approach creates frequent false negatives: when a modern embedding model retrieves a genuinely relevant document that human annotators missed years ago, legacy benchmarks treat the result as a failure.

To fix that distortion, Cohere researchers Fabian David Schmidt, Donato Crisostomi, Carlos Lassance, and Nils Reimers built RCP-nDCG@10 (Rubric-Calibrated Preferences nDCG@10). The method replaces static labels with query-specific qualitative rubrics evaluated by language models and calibrated statistically through Item Response Theory (IRT). The team published its theoretical foundation on arXiv (paper 2609.35739) and open-sourced the code on GitHub. Kenneth Enevoldsen, lead maintainer of MTEB at Aarhus University, publicly praised the framework and expressed interest in integrating it into the community benchmark.

In-house numbers and necessary caveats

In internal tests run on the ViDoRe V3 document benchmark using RCP-nDCG@10, Cohere reported scores of 85.8 for Embed 5 Pro and 84.5 for Embed 5 Fast. The company pointed to those marks to claim an advantage over rivals, citing Voyage 4 Large at 83.7, Gemini Embedding 2 at 83.2, and OpenAI’s text-embedding-3-large at 75.5. In a company validation test with 46 contracted human reviewers, the new method showed 70 percent agreement across 289 dispute cases.

Those results warrant clear technical skepticism. Building both the model and the yardstick used to evaluate it creates an obvious conflict of interest. Any claim of superior performance verified by neutral third parties outside Cohere remains RUMOR. Cohere itself notes that Embed 5 was tuned specifically for this calibrated methodology and might score lower under legacy fixed-label metrics. Furthermore, reports that RCP-nDCG@10 is already live on MTEB’s public Hugging Face leaderboard are RUMOR. The real-world value of Embed 5 will depend on how the models perform inside actual production databases.