Cohere Command A+ Review 2026: The Open-Weight Enterprise Agent Model That Does More for Less
Cohere Command A+ Review 2026: The Open-Weight Enterprise Agent Model That Does More for Less
Data verified: September 1, 2026. Prices follow Cohere's listed API rates and are subject to change.
While everyone obsesses over trillion-parameter frontier models, Cohere quietly shipped one of the most practically useful AI models of 2026: Command A+ (command-a-plus-05-2026), released on May 20, 2026 under the permissive Apache 2.0 license. It is a 218-billion-parameter sparse mixture-of-experts model with only 25B active parameters per token — open weights, enterprise-grade tool use, vision input, 48 languages, and a price tag that undercuts closed frontier APIs by an order of magnitude.
In this review we separate the measurable reality from the marketing: what Command A+ is, how it benchmarks against its predecessor and the competition, what it costs to run, where it genuinely shines for real enterprise workloads, and where it still falls short.
The Verdict in One Sentence
Command A+ is the best open-weights "do it yourself" enterprise agent model in 2026 — a genuinely open (Apache 2.0) 218B MoE that triples agentic-tool performance, reads images, speaks 48 languages, and runs on two H100s at 4-bit — but its 128K context and April-2025 knowledge cutoff mean it is a pragmatic workhorse, not a reasoning-fantasy frontier model.
Core Specs at a Glance
| Spec | Command A+ (2026) |
|---|---|
| Release date | May 20, 2026 |
| Total / active parameters | 218B / 25B (sparse MoE, 128 experts, 8 active + 1 shared) |
| Architecture | Decoder-only sparse MoE, sliding-window attention + RoPE, 3:1 interleaved global attention |
| Context window | 128K input / 64K max output |
| Knowledge cutoff | April 1, 2025 |
| Modalities | Text + image input; text, reasoning, tool-use output |
| Languages | 48 (all official EU languages) — up from 23 |
| License | Apache 2.0 (open weights: BF16 / FP8 / W4A4 on Hugging Face) |
| Min hardware | W4A4: 1× B200 or 2× H100 |
| Frameworks | vLLM (≥0.21.0), Transformers + Cohere "melody" library |
Pricing: Same Price, Far More Model
Command A+ costs exactly the same as the original Command A on Cohere's API, making it a price-neutral upgrade:
| Token type | Price per 1M tokens |
|---|---|
| Input | $2.50 |
| Output | $10.00 |
| Cache read | $1.25 |
| Trial / production keys | Free until rate limits are reached |
For context, that is roughly 2.5× cheaper than a comparable closed frontier agent API on blended pricing (e.g. Claude Opus 4.7 lists at $25/$125) and only modestly above efficient open-weights peers like Mistral Large 3 ($1.80/$5.40) and Llama 4. And because the weights are Apache 2.0, the API price is only one option — you can self-host and pay for GPUs instead of tokens.
Benchmarks: Where Command A+ Actually Landed
Cohere's headline story is the jump in agentic performance, and the numbers back it up. Gains shown are versus Command A Reasoning unless noted:
| Benchmark | Command A+ | Command A Reasoning |
|---|---|---|
| τ²-Bench Telecom (agentic tool use) | 85% | 37% |
| Terminal-Bench Hard (agentic coding) | 25% | 3% |
| AIME 2025 (math) | 90% | — |
| GPQA Diamond (science reasoning) | ~76% | — |
| MMMU (multimodal understanding) | 75.1% | 65.3% |
| MMMU Pro | 63% | — |
| MathVista | 80.6% | 73.5% |
| CharXiv reasoning (chart understanding) | 52.7% | 46.9% |
| Artificial Analysis Intelligence Index | 37 | — |
| AA-Omniscience Non-Hallucination | 86% (rank #1) | — |
The τ²-Bench Telecom jump from 37% to 85% is the single most meaningful number here. It signals that Command A+ was trained specifically to be reliable when given a live API, database or search endpoint and asked to get a task done end-to-end — which is precisely what enterprise agents burn out on.
Its Artificial Analysis Intelligence Index of 37 places it around the level of Claude 4.5 Haiku — solidly mid-to-upper tier, clearly behind the current reasoning kings (Opus 4.8-class, GPT-5.5-class) but dramatically ahead of models at its price and size class.
Efficiency: The Second Headline
Command A+ was built to run cheaply:
- Up to 110% higher throughput and 30% lower latency than Command A Reasoning.
- W4A4 (4-bit) quantization gives 47% faster inference and 13% lower latency with negligible quality loss — and it is the recommended format.
- Speculative decoding delivers a further 1.5–1.6× speedup.
- A new multilingual tokenizer cuts token counts by ~20% for Arabic, 16% for Korean and 18% for Japanese.
- Roughly 281 output tokens/second on Cohere's API — faster than GPT-5.4 nano, Claude 4.5 Haiku and Grok 4.3, just behind Gemini 3.1 Flash-Lite (304 t/s).
That efficiency is why Cohere can claim enterprise-grade agents on commodity hardware: two H100s at W4A4 is a realistic on-prem footprint for a 218B-parameter model that outperforms its predecessors on everything that matters to agents.
Enterprise Story: North, Model Vault and Sovereign AI
Command A+ is not a consumer chatbot — it is the engine Cohere built from a year of running North, its enterprise agent workspace. Internal North evaluations show a +20% jump in agentic QA accuracy and +32% better spreadsheet analysis versus Command A Reasoning, with long-term memory performance improving from 39% to 54%.
Three things make it enterprise-grade out of the box:
- Native citations — grounded, structured source links in answers, essential for compliance-heavy RAG workflows.
- Structured outputs and safety modes — predictable JSON and configurable guardrails for production systems.
- Model Vault — managed, isolated inference for customers who want the safety of a hosted API without sharing infrastructure.
It has already been endorsed by Fujitsu for sovereign AI offerings (its Takane model and Kozuchi Enterprise AI Factory), which is a strong signal for public-sector and regulated-industry buyers in Europe and Japan.
Open Weights: The Real Differentiator
Apache 2.0 is the important word. Unlike almost every frontier model, Command A+ gives you:
- Full ownership of your data — no API, no logging of your prompts, no vendor lock-in on inference.
- Air-gapped / on-prem deployment for banks, healthcare and government.
- Fine-tuning freedom — it is your model to adapt.
- Commercial use without royalties.
This puts it in the same open-weights class as Llama 4, DeepSeek, Qwen and Mistral — with a genuinely different emphasis: Cohere tuned it for agents that use tools, not just chat. The trade-off is a smaller surrounding ecosystem than Meta's or Alibaba's — fewer community fine-tunes, less third-party tooling — and a dependency on Cohere's melody library for W4A4 inference.
Scorecard
| Dimension | Score | Note |
|---|---|---|
| Agentic / tool use | 9.5/10 | 85% τ²-Bench Telecom; built for production agents |
| Openness & licensing | 10/10 | Apache 2.0, fully self-hostable |
| Cost efficiency | 9/10 | $2.50/$10 API; runs on 2× H100 at W4A4 |
| Multilingual & multimodal | 8.5/10 | 48 languages, strong vision (MMMU 75.1%) |
| Raw reasoning | 7.5/10 | Solid (AA index 37) but not frontier-class |
| Context length | 6.5/10 | 128K is fine, but rivals now offer 1M–2M |
| Ecosystem / tooling | 7/10 | vLLM/Transformers OK; community smaller than Llama/Qwen |
Pros & Cons
What Command A+ does really well
- Genuinely open: Apache 2.0 weights in three quantizations — deploy anywhere, keep your data.
- Leaps ahead on agents: τ²-Bench Telecom 85% and Terminal-Bench Hard 25% show it was purpose-trained for real tool use.
- Cheap to run: $2.50/$10 API pricing and a realistic 2× H100 on-prem footprint at 4-bit.
- One model, many jobs: reasoning, vision, translation and tool use unified — no juggling separate models.
- Lowest hallucination rate on AA-Omniscience (86%), a critical property for enterprise RAG.
- Truly multilingual: 48 languages including every official EU language.
Where it falls short
- 128K input context — smaller than its own predecessor's 256K and far behind Qwen/Kimi/Gemini million-token models; long-document teams will notice.
- Knowledge cutoff April 2025 — over a year old by late 2026; it will hallucinate on recent events without RAG.
- Not a reasoning king: AA Intelligence Index 37 sits at Haiku-class, not Opus/GPT-5.5-class — don't expect frontier math or codegen.
- Smaller ecosystem: fewer community fine-tunes and third-party tools than Llama 4 or Qwen; W4A4 needs Cohere's
melodylibrary. - Enterprise-first positioning — no consumer app, limited developer community around it.
Frequently Asked Questions
Is Command A+ better than the original Command A?
Overwhelmingly yes for agents: τ²-Bench Telecom 37% → 85%, Terminal-Bench Hard 3% → 25%, MMMU 65.3% → 75.1%, plus vision, reasoning, 48 languages and up to 110% higher throughput — at the same $2.50/$10 API price. The main regression is context: 128K input versus Command A's 256K.
Can I run Command A+ on my own GPU?
Yes. At W4A4 you need a single B200 or two H100s. FP8 needs 4 H100s and BF16 needs 8. It works with vLLM ≥0.21.0 and Transformers; use Cohere's melody library for the 4-bit format.
Is Command A+ good for non-English content?
Excellent. It covers 48 languages including every official EU language, and the new tokenizer is ~16–20% more efficient for Arabic, Korean and Japanese — a genuinely strong choice for multilingual enterprise RAG and customer support.
Why would I choose Command A+ over Llama 4 or Qwen?
Both are fine Apache/MIT open-weights alternatives. Choose Command A+ if your workload is agents using tools (APIs, databases, search) and you want the strongest turnkey tool-use + citations behavior, plus best-in-class hallucination control. Choose Llama/Qwen for a bigger ecosystem, larger context, or if you want the cheapest possible massive-context open model.
Does Command A+ have a consumer chatbot?
No. Cohere sells to developers and enterprises (API, Model Vault, North). Consumers should use one of the many chat apps that route through Cohere instead.
Bottom Line
Cohere Command A+ is the open-weights model I would actually deploy for an enterprise agent workload in 2026. It is open (Apache 2.0), fast (2× H100 at 4-bit), cheap ($2.50/$10), multilingual, low-hallucination, and — crucially — it was trained for the thing enterprise teams actually struggle with: getting an AI to reliably use tools end-to-end.
It is not the strongest raw-reasoning model in the world, and its 128K context and April-2025 knowledge cutoff will bother teams that live on million-token documents. But as a sovereign, privately deployable, genuinely open agent engine, nothing at its price point does more for less.
Score: 8.3/10 — Recommended for enterprise agent and RAG teams; look elsewhere only if you need million-token context or absolute top-tier reasoning.
评论
发表评论