Command A+
Released May 20, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- Canada
- Modality
- Multimodal
- Context window
- 128K
- Max output
- 64K
- Speed
- 193 tok/s · 0.44s to first answer token
- Price ($/MTok in / out)
- — / —
- 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
- Cohere's first Mixture-of-Experts release (218B total / 25B active), fitting on a single B200 or two H100s despite consolidating the whole Command A generation into one model
- Large jump on Cohere's own agentic-coding and telecom-agent suites versus its immediate predecessor — Terminal-Bench Hard 3% to 25% and tau-squared-Bench Telecom 37% to 85%
- Adds vision input and expands language coverage from 23 to 48 languages while unifying what used to be four separate Command A variants (base, Reasoning, Vision, Translate) into one model
- Apache 2.0 weights make it commercially self-hostable, unlike Command A's non-commercial CC-BY-NC licence
Weaknesses
- Released without a standard cross-lab benchmark table (no MMLU-Pro, GPQA, SWE-bench, AIME or ARC-AGI-2 figures published) — Cohere's own comparisons are against its prior Command A generation and Cohere-specific/multimodal suites (MMMU, MathVista, CharXiv), not external frontier baselines
- Not listed on arena.ai's Text leaderboard or any independent tracker's benchmark suite as of this check, so head-to-head positioning against non-Cohere models is largely unverifiable
- Too new for hands-on developer reception at time of research — behavior in production agent loops is not yet documented outside Cohere's own launch material
Cohere's launch post (cohere.com/blog/command-a-plus) states it 'surpasses every previous generation in the Command series and unifies their capabilities into a single scalable model,' explicitly naming Command A among what it consolidates — read as an explicit supersession statement, so predecessorId is set to command-a. Reported figures (MMMU Pro 63%, MMMU 75.1%, MathVista 80.6%, CharXiv 52.7%, tau-squared-Bench Telecom 85%, Terminal-Bench Hard 25%, Artificial Analysis Intelligence Index 37) are Cohere's own and don't map onto this site's tracked eight benchmark keys, so all eight are left null; Terminal-Bench Hard is a different suite from the tracked Terminal-Bench 2.1. Pricing is unpublished as of this check — Cohere's pricing page and docs both defer newer Command A models to a sales-contact flow, and Artificial Analysis shows $0.00/$0.00 for its tracked build, which reads as a data placeholder rather than a real price and was not used. Context window conflict: Cohere's own blog and docs state 128K input / 64K max output; Artificial Analysis instead lists a 192K context window for its tracked variant — used Cohere's own figure and flagged the conflict rather than picking silently. Speed (193.4 tok/s, 0.44s TTFT) is from Artificial Analysis' page for what it labels the model's 'reasoning version'; Cohere describes reasoning as a supported mode without publishing distinct low/medium/high effort tiers, so effort is left null rather than guessed. Apache-2.0 licence and 218B-total/25B-active parameter count are from the Hugging Face model card. Knowledge cutoff not disclosed.
For developers
API model strings
- cohere
command-a-plus-05-2026
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
Replaces Command A