Open SourceCohereReleased 2026-05

Command A+

Cohere's first mixture-of-experts Command model: open weights, vision input and reasoning on two H100s

Context window

128K

Input / 1M tokens

$0.30

Output / 1M tokens

$1.50

Provider

Cohere

Cohere's pricing page does not list a per-token price for Command A+. The price shown is OpenRouter's (cohere/command-a-plus, served by Cohere), which also lists cached input at $0.15 per 1M. Open weights under Apache 2.0 are free to self-host, and Cohere Model Vault offers dedicated managed inference billed per GPU hour. · Data verified 2026-11-11

Command A+ (May 20, 2026) is Cohere's newest flagship in the Command A family and its first mixture-of-experts model, with 218B total and 25B active parameters. It brings agentic, reasoning, vision and translation abilities into one model, takes text and image input with a 128K context and up to 64K output tokens, and supports 48 languages including all official EU languages. Cohere released the weights under Apache 2.0 in BF16, FP8 and W4A4; the W4A4 version runs on one B200 or two H100 GPUs. Cohere reports large gains over Command A Reasoning, such as 85% vs 37% on tau2-Bench Telecom.

How to access it

Cohere API (model ID command-a-plus-05-2026, listed as Live on Cohere's models page), OpenRouter (cohere/command-a-plus), Cohere Model Vault for dedicated inference, and open weights on Hugging Face (CohereLabs, in BF16, FP8 and W4A4) for self-hosting with vLLM or Transformers.

Strengths

  • ✓Apache 2.0 open weights that run on one B200 or two H100s at W4A4
  • ✓Agentic tool use: 85% on tau2-Bench Telecom (Cohere-reported)
  • ✓Image input with reasoning, plus support for 48 languages
  • ✓Up to 63% higher output speed than Command A Reasoning (Cohere-reported)

Best for developers who...

Private or on-premises enterprise agentsMultilingual assistants and translationDocument and chart understanding with imagesRAG and tool-use workflows

When to choose it (and when not to)

Reach for Command A+ when...

  • →You want an open-weight enterprise model you can run privately on two H100s
  • →You need multilingual agents across 48 languages
  • →You run an older Command model and want Cohere's current flagship

Look elsewhere if...

  • ✕You need top frontier coding scores
  • ✕You need more than 128K tokens of context
  • ✕You need a per-token price published by Cohere itself

How to use it

  • ›Turn thinking on for multi-step agent tasks and off for quick replies, and cap its token budget
  • ›Pass source documents to the Chat API to get grounded answers with citations
  • ›Use the W4A4 weights when self-hosting on limited GPUs

Quickstart

Python
import cohere

co = cohere.ClientV2(api_key="YOUR_COHERE_API_KEY")

resp = co.chat(
    model="command-a-plus-05-2026",
    messages=[{"role": "user", "content": "Summarize the main risks in this contract: ..."}],
)
for block in resp.message.content:
    if block.type == "text":
        print(block.text)

Requires the official 'cohere' Python SDK (pip install cohere). The loop skips thinking blocks when reasoning is on.

API model id: command-a-plus-05-2026

Benchmarks

BenchmarkScoreNotes
tau2-Bench Telecom85%Cohere-reported; Command A Reasoning 37%
Terminal-Bench Hard25%Cohere-reported; Command A Reasoning 3%
MMMU75.1%Cohere-reported; Command A Vision 65.3%
MMMU Pro63%Cohere-reported
MathVista80.6%Cohere-reported; previous model 73.5%
Artificial Analysis Intelligence Index37Third-party, cited in Cohere's launch post

Source: Command A+ announcement (Cohere blog)

Compare Command A+

Compare Command A+ with any other model

Build a comparison →
All model comparisons →

Learn the concepts