Wait Which Model?
← Back to directory
Cohere

Command A

Released Mar 13, 2025 · knowledge cutoff 2024-08

Status
Superseded
Location
Canada
Modality
Text
Context window
256K
Max output
8K
Speed
60 tok/s · 1.71s to first answer token
Price ($/MTok in / out)
$2.50 / $10
Cost per task
—
Open weights
Yes
Benchmarks0 of 8 reported

MMLU-Pro

—

GPQA Diamond

—

SWE-bench Verified

—

Terminal-Bench 2.1

—

AIME (latest)

—

Humanity's Last Exam

—

LMArena Elo

—

ARC-AGI-2

—
Strengths & weaknesses

Strengths

  • Runs on two GPUs rather than the eight-plus a comparable 100B+-class model typically needs, making it one of the cheaper flagships to self-host at the 111B scale
  • Restrained tool-calling — Cohere's own docs specifically credit it with avoiding unnecessary tool calls, a common failure mode for agent-oriented models
  • Strong at attending over long financial/numerical inputs and RAG-style grounded generation across its 23 supported languages

Weaknesses

  • Chatty by default — Cohere's own documentation warns it's verbose and markdown-heavy out of the box, and recommends system-prompt overrides for concise output
  • No published MMLU-Pro/GPQA/SWE-bench-class benchmark table at launch, unusual for a same-tier 2025 flagship — Cohere's headline comparisons (vs. GPT-4o, DeepSeek-V3) are qualitative rather than tabulated scores
  • Weights are CC-BY-NC only, so open-weight users can inspect and fine-tune but not deploy commercially without a separate Cohere licence

Cohere's own docs (docs.cohere.com/docs/command-a) state '150% higher throughput compared to its predecessor, Command R+ 08-2024' — an explicit predecessor claim, but Command R+ isn't a tracked model in this dataset, so predecessorId stays null (logged to spec-gaps.md). No independent or official numeric benchmark table (MMLU-Pro/GPQA/SWE-bench/AIME/HLE/ARC-AGI-2) was found — Cohere's launch blog renders its comparison charts as images rather than text, and Command A isn't listed on arena.ai's Text leaderboard; all eight tracked fields left null rather than approximated from the qualitative 'on par with GPT-4o and DeepSeek-V3' claim. Pricing ($2.50/$10.00) and speed are corroborated across OpenRouter and Artificial Analysis (artificialanalysis.ai/models/command-a); Cohere's own pricing page did not list current Command A/A+ rates when checked, directing to the (paywalled/enterprise) sales channel instead. Knowledge cutoff (Aug 2024) is from OpenRouter's model listing, not confirmed on an official Cohere page — treat as third-party-sourced. License is CC-BY-NC-4.0 plus Cohere's Acceptable Use Policy per the Hugging Face model card (huggingface.co/CohereLabs/c4ai-command-a-03-2025); 111B total parameters.

For developers

API model strings

  • coherecommand-a-03-2025

Licence

CC-BY-NC 4.0 + Cohere Acceptable Use Policy · Restricted · commercial use restricted

Retirement

No retirement announced

Lineage

No recorded predecessor

News