Cohere
Cohere’s multimodal embedding model for text, images, and mixed documents, with 128K context and configurable embedding dimensions for enterprise search/RAG.
embed-v4.0 is a specialized embedding API rather than a general-purpose generative or agentic model. Official Cohere material supports text, image, and mixed-document embeddings, a 128K context window, and configurable dimensions. Its multimodal I/O score is therefore stronger than GPT-5.6 Terra and GPT-5.6 Sol, whose published anchors have narrower stated multimodal scores, while remaining below Gemini 3.5 Flash’s 92 because no independent comparative retrieval or vision-embedding result was supplied. Technical and developer scores reflect a documented, purpose-built retrieval offering, not demonstrated frontier general intelligence. Coding, agentic work, and reasoning scores are materially below Bria 3.2 and the GPT, Claude, Gemini, and Kimi anchors: DeepSWE and LiveCodeBench list no result, SWE-bench and Terminal-Bench/Aider evidence was not found, and these tasks are outside the model’s embedding role. The same scope limitation makes LMArena/Arena-Hard preference rankings inapplicable. Cohere’s documentation and API positioning support solid developer experience for enterprise RAG, but the supplied material provides no verified latency, availability, throughput, or independent quality comparison. A Cohere pricing page was checked, yet supplied evidence contains no embed-v4.0 rate details, holding down pricing clarity and cost-effectiveness. Evidence quality is moderate: official claims are clear, but public third-party benchmark coverage and direct adoption signals are limited.
