Command A+ is Cohere's successor to Command A, released May 20, 2026 as an open-weight model under Apache 2.0. It's a 218B-parameter mixture-of-experts model with 25B active parameters that takes text and image input, and Cohere doesn't publish a metered per-token price for it.
There's no public per-token price for Command A+: Cohere says it's available to all users on its standard API, but its pricing page lists no rate for it, and free trial keys can't be used commercially. For production, Cohere offers Model Vault (priced per instance), private deployments and Azure AI Foundry. Because the weights are Apache 2.0, the practical price for many teams is the GPU cost of hosting it: Cohere says it fits on one B200 or two H100s at W4A4 quantization.
| Command A+ | |
|---|---|
| Provider | Cohere |
| Tier | Specialist |
| Input / 1M tokens | — |
| Output / 1M tokens | — |
| Cache hit / 1M | — |
| Cache write / 1M | — |
| Context window | 128K |
| Notes | Succeeds Command A; open-weight (Apache 2.0); live on Cohere's standard API (command-a-plus-05-2026) and Azure AI Foundry, but Cohere publishes no per-token rate; 64K max output |
| Model | Tier | Input / 1M | Output / 1M | Cache hit / 1M | Context | Notes |
|---|---|---|---|---|---|---|
| Codestral | Specialist | $0.30 | $0.90 | $0.03 | 128K | Code generation |
| Grok Build 0.1 | Specialist | $1.00 | $2.00 | $0.20 | 256K | Code tasks; replaced Grok Code Fast 1; ≥200K tokens: $2/$4, cache $0.40 |
| Sonar | Specialist | $0.25 | $2.50 | $0.0625 | 128K | Cheap web-grounded answers; now served via the Agent API (perplexity/sonar) after Sonar Chat Completions support ended Sep 27, 2026; billed per token plus per tool call (web_search $0.0025 per invocation, $0.001 with Fast Search) |
You can reach Command A+ (command-a-plus-05-2026) several ways. Cohere's changelog says it's available to all Cohere users through the standard Chat endpoint. Trial API keys are free but rate-limited and can't be used for production or commercial work. For production, Cohere offers Model Vault, a dedicated managed inference environment priced per instance rather than per token, and private deployments for enterprises. It's also on Azure AI Foundry as coherelabs-command-a-plus-05-2026-w4a4. Finally, the weights are on Hugging Face under Apache 2.0, so you can self-host with no licence fee.
Only 25B of the 218B parameters are active per token, which keeps inference compute low for its size, but the full model still has to fit in GPU memory. Cohere lists the minimum as one B200 or two H100s at W4A4 quantization, with BF16 and FP8 weights also published; W4A4 needs vLLM 0.25.0 or later. Price the GPU-hours you'd reserve, divide by the tokens you expect to serve, and compare the result with the per-token rates of the hosted model you'd otherwise use.
Cohere positions Command A+ as one set of weights combining vision input, agentic use, reasoning and translation, and it supports 48 languages, including all official EU languages. The case is strongest when data residency or air-gapped deployment rules out third-party APIs. If you just want a hosted model billed per token, a general-purpose API with published rates is simpler to start with; the specialist models listed above are code and search models, not like-for-like substitutes.
Cohere doesn't publish a metered per-token price for Command A+. Cohere says it's on the standard API for all users, but its pricing page lists no rate. Trial keys are free but rate-limited and non-commercial; for production, Cohere offers Model Vault (per-instance pricing) and private deployments, it's also on Azure AI Foundry, or you can self-host the open weights.
Yes. Cohere released Command A+ under the Apache 2.0 license, with weights on Hugging Face in BF16, FP8 and W4A4 formats.
Cohere lists a minimum of one NVIDIA B200 or two H100 GPUs at W4A4 quantization.
Yes. It accepts text and image input and produces text output.
How we choose: prices come from each vendor's official pricing page and are re-checked every week. We take no affiliate fees, so rankings aren't for sale.