command-r-plus-08-2024

Command R+ is Cohere's flagship model, optimized for retrieval-augmented generation (RAG), enterprise search, and tool use. It is specifically designed for production RAG pipelines, with native grounding capabilities that reduce hallucination when answering questions over retrieved documents. It supports a 128K-token context window, tool use, and streaming, and is available via Cohere's API and major cloud providers.

Command R+ is priced at $2.50 per million input tokens and $10.00 per million output tokens. At 1,000 input + 500 output tokens per call, it costs approximately $0.0075 per 1,000 API calls. For RAG-heavy workloads, the pricing is competitive given Command R+'s specialized capability in this area. Cohere also offers a "Command R" model at $0.15/$0.60 per million tokens for less demanding RAG tasks, making the Command family a cost-efficient option for RAG pipelines at different quality tiers.

Command R+'s 128K-token context window is suitable for most RAG use cases, enabling retrieval of many document chunks in a single call. Median latency is approximately 400–600ms TTFT. Throughput is approximately 70–90 tokens per second. Maximum output is 4,096 tokens — lower than GPT-4o, which may be a constraint for applications requiring long generated responses. The model's grounding capability — citing specific retrieved passages in its responses — is a native feature rather than a prompt engineering workaround, making it more reliable for enterprise document Q&A applications.

Command R+ is best suited to enterprise RAG pipelines, document Q&A, enterprise search, and any application where grounding responses in retrieved documents is important. Its native citation capability reduces hallucination in RAG settings — the model is trained to distinguish between what it knows from training and what it has retrieved, and to cite sources explicitly. Tool use is reliable for structured data extraction and API integration. For teams building production RAG systems, Command R+ is worth evaluating — its specialized training for RAG tasks can deliver better grounding quality than general-purpose models at comparable cost.

Command R+ sits at the top of Cohere's model lineup, above Command R (smaller, cheaper) and Command Light (fastest, cheapest). Its main competitors for RAG use cases are general-purpose frontier models from OpenAI and Anthropic, and open-weights models that can be self-hosted but lack native grounding. For teams specifically building RAG applications, Command R+'s specialized training is its key differentiator. For general-purpose applications, other advanced-tier models offer broader capability. Its intelligence score of approximately 76/100 reflects strong RAG-specialized performance rather than frontier general capability.