Cohere's enterprise text model for tool-using agents and cited RAG, with a 256K context window and 23-language coverage.
Modality:
Text
model ID
cohere/command-a
Output Speed *
53.42
tok/s
Intelligence Index *
7
/ 100
Context Window *
256000
tokens
Input price
15
Anytoken
Output price
60
Anytoken
Command A: Enterprise Agents and Cited RAG at 256K Context
Command A is Cohere's flagship text model for enterprise workloads. It is a 111B-parameter model built for tool use, retrieval-augmented generation with citations, and multilingual work. It succeeds Command R+ and sits above Command R7B in the Cohere lineup. Cohere designed it to run on two GPUs when self-hosted. It suits agentic workflows, grounded question answering over large document sets, and multilingual assistants that need a long context window but not image input or an explicit reasoning mode.
Access Command A via the AnyAPI.ai API
Performance
Where Command A Earns Its Place: Grounded, Tool-Driven Enterprise Work
Command A is tuned for the tasks enterprise teams actually run: tool calling, retrieval-grounded answers with citations, and multilingual generation. Cohere reports a throughput of about 156 tokens per second at launch and roughly 150% higher throughput than Command R+ (08-2024). These are provider-reported numbers, not independent measurements. The practical consequence is that agent loops and RAG pipelines can use a long 256K context without a very large GPU footprint. Treat it as a workhorse for grounded text tasks, not a frontier reasoning model.
Technical Specifications
What the model supports
Command A accepts and returns text only. It has a 256K-token context window but a maximum output of roughly 8K tokens, so it can read very large inputs but cannot write book-length responses in one call. This matters most for document-heavy RAG: plan retrieval and chunking around the 8K output cap. The model has no built-in reasoning mode. Cohere offers separate Command A variants for vision, translation and reasoning.
Verified Specifications —
Command A
*
Input modalities
Text
output modalities
Text
Context window
256000
tokens
Maximum output tokens
8192
Reasoning
No
Knowledge cutoff
March 2025
Pricing (standard)
15
AnyTokens in
/
60
AnyTokens out
Limitations & Trade-offs
1
Text-only input. Command A accepts and produces text only, so screenshots, scanned pages and charts need an OCR or captioning step first. Cohere ships Command A Vision as a separate model. If your pipeline processes documents as images, or needs chart and diagram understanding, use a vision-capable model instead of adding preprocessing in front of this one.
2
Output cap far below context. The context window is 256K tokens, but maximum output is about 8K. Tasks that read a huge corpus and must return a long report, full codebase rewrite or large translation in one response will hit the cap. You must split generation into several calls. Models with larger output limits fit long single-pass generation better.
3
No reasoning mode. Command A has no extended-thinking mode and no reasoning-effort controls. It handles tool routing and grounded answers well, but hard multi-step math, formal logic or deep planning are better served by a dedicated reasoning model. Cohere offers Command A Reasoning separately. Any other reasoning-class model is a better fit for those problems.
4
Non-commercial weights license. Cohere has published Command A weights for research under a non-commercial license. Self-hosting for production commercial use therefore needs a separate agreement with Cohere. Teams that want to avoid that licensing and GPU operations work will find API access the simpler route, and should confirm licensing terms before planning on-premises deployment.
Best-Fit Workloads
Where this model earns its place
01
Cited RAG over large knowledge bases
Command A's 256K context window lets you pass many retrieved passages in one request. Cohere positions the model around retrieval-augmented generation with citations, which suits internal search, policy Q&A and support assistants that must show sources. The 8K output cap fits answer-style responses. It is a weaker choice if the pipeline relies on images of documents, which the model cannot read.02
Tool-using enterprise agents
The model supports tool calling and JSON-schema-constrained outputs, and Cohere designed it for agentic enterprise tasks. That makes it suitable for agents that query CRMs, ticketing systems or databases and return structured results. For long, deeply planned multi-step tasks that need explicit reasoning, add a reasoning model in the planning step.03
Multilingual customer-facing assistants
Command A covers 23 languages, including Arabic, Japanese, Korean, Hindi and several European languages. Combined with citation-backed answers, it suits support and internal-help assistants that serve several regions from one model. Quality still varies by language, so evaluate on your own target languages before launch.04
Structured extraction from long text
With structured outputs and a 256K window, Command A can turn long contracts, reports or transcripts into schema-conforming JSON. Extraction outputs are usually short, so the 8K output limit rarely matters. Documents that exist only as scans or images need OCR first, because the model is text-only.Integration
01
One API integration
02
Easy model switching
03
Flexible for production
04