Cohere Command A is a family of five enterprise large language models from Cohere, starting with Command A (command-a-03-2025) and ending with Command A+ (command-a-plus-05-2026), which Cohere's changelog calls "the last model in the Command A family". This profile collects the published facts about all five members in one place: API identifiers, parameter counts, context windows, output limits, knowledge cutoffs, prices, hardware needs, availability channels, licences, data handling and lifecycle status. Models are listed by release date, newest first, an objective order that implies no ranking. Every fact comes from Cohere's own documentation, pricing page, changelog, data commitments page and Cohere Labs model cards, listed under Standards and references. Where Cohere publishes nothing on a point, the page says so.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Cohere's public documentation as of November 2026. Model details change, so confirm against Cohere's current documentation.
Key takeaways
- Five models, all live. Cohere's models page lists Command A+, Command A Reasoning, Command A Translate, Command A Vision and Command A with status Live, and none of them appears on the deprecation notices as of 15 November 2026.
- Command A+ is, per Cohere's changelog, its first Mixture of Experts model. Released on 20 May 2026, it has 218B total and 25B active parameters, takes text and image input, supports 48 languages and runs on one B200 or two H100 GPUs at W4A4 precision.
- Context runs from 8K to 256K tokens. Command A and Command A Reasoning list 256K, Command A+ and Command A Vision list 128K and Command A Translate lists 8K; maximum output ranges from 8K to 64K.
- Token prices are published for Command A only. The Command A model page lists $2.50 input and $10 output per million tokens; the other four pages say they are free on trial and production keys until rate limits are reached, with production use through Model Vault or Cohere sales.
- Two members have published weights. Cohere Labs publishes Command A+ under Apache 2.0 and Command A under CC-BY-NC 4.0 with an Acceptable Use Policy; a licence for the other three is not stated on the pages checked.
What is Cohere Command A?
Command A is the generation of Cohere's Command chat models that followed Command R and Command R+. The Command A model page describes the first member as strong at tool use, agents, retrieval-augmented generation (RAG) and multilingual work, with 150% higher inference throughput than Command R+ 08-2024 and a footprint of two GPUs. Cohere then released specialised variants built on the same 111B-parameter size: a reasoning model, a vision model and a translation model. In May 2026 Command A+ combined those capabilities in one set of weights on a new sparse architecture.
By release order the members are Command A (March 2025), Command A Vision (July 2025), Command A Reasoning and Command A Translate (August 2025) and Command A+ (May 2026); the comparison table near the end of this page sets out how they differ.
All five are served through Cohere's Chat endpoint (Chat V2, Chat V1 and the OpenAI-compatible Chat Completions endpoint) and the Playground. For the wider Cohere line-up, including North, Embed, Rerank and Aya, see the Cohere Models hub.
Key facts at a glance
As of 2026-11-15. Source: Cohere models page, individual model pages, changelog and Cohere Labs model cards. Values are listed per model, newest first.
| Field | Value | Source |
|---|---|---|
| Provider | Cohere Inc. | Command A+ model page |
| API IDs | command-a-plus-05-2026; command-a-translate-08-2025; command-a-reasoning-08-2025; command-a-vision-07-2025; command-a-03-2025 | Models page |
| Release date | Command A+: 2026-05-20; other members: month and year from the ID suffix (08-2025, 08-2025, 07-2025, 03-2025) | Command A+ model page; changelog; models page |
| Status | Live (all five) | Models page |
| Context window | A+ 128K; Translate 8K; Reasoning 256K; Vision 128K; A 256K | Models page; model pages |
| Max output | A+ 64K; Translate 8K; Reasoning 32K; Vision 8K; A 8K | Models page; model pages |
| Knowledge cutoff | A+: 1 April 2025; Translate, Reasoning, Vision and A: 1 June 2024 | Model pages |
| Input modalities | A+ and Vision: text, image; Translate, Reasoning and A: text | Models page |
| Output modalities | Text (all five) | Models page; model cards |
| Parameters | A+: 218B total, 25B active (MoE); Translate, Reasoning and A: 111B; Vision: Not stated by Cohere on the pages checked (2026-11-15) | Model pages; changelog |
| Open weights | A+: yes (Apache 2.0); A: yes (CC-BY-NC 4.0); Translate, Reasoning, Vision: Not stated by Cohere on the pages checked (2026-11-15) | Cohere Labs model cards; Command A+ model page |
| Languages | A+: 48, including all official EU languages; Translate, Reasoning and A: 23; Vision: English, Portuguese, Italian, French, German and Spanish officially supported | Model pages; models page |
The Hugging Face header for Command A+ shows 219B parameters, while the card text and Cohere's changelog state 218B total. The Command A card says the Hugging Face configuration is set to 128K context by default and can be changed to use the full 256K.
Release and version history
As of 2026-11-15. Dates from the Cohere changelog and model pages; where Cohere gives no exact date, the month comes from the model ID.
| Date | Event | Source |
|---|---|---|
| 2025-03 (from ID) | Command A (command-a-03-2025) released; the Command A Technical Report is linked from its model page and the linked paper on the model card is dated 1 April 2025 | Command A model page; Command A model card |
| 2025-07 (from ID) | Command A Vision (command-a-vision-07-2025) | Models page |
| 2025-08 (from ID) | Command A Reasoning (command-a-reasoning-08-2025) and Command A Translate (command-a-translate-08-2025) | Models page |
| 2025-09-15 | Original Command, Command Light, Command R (03-2024) and Command R+ (04-2024) deprecated, leaving Command A and the 08-2024 R models as the Command line | Models page |
| 2026-04-04 | Embed v2.0 and Aya Expanse and Aya Vision 8B retired; command-a-03-2025 and command-a-reasoning-08-2025 named among recommended chat replacements | Changelog |
| 2026-05-20 | Command A+ announced as the last model in the Command A family, with Apache 2.0 weights and private deployment for enterprise customers | Changelog; Command A+ model page |
Each model has one dated ID, and Cohere lists no later snapshots under the same names.
How much does Cohere Command A cost?
As of 2026-11-15. USD per million tokens where a token price is stated. Source: Cohere model pages and pricing page.
| Item | Price | Notes | Source |
|---|---|---|---|
| Command A+ tokens | Free on trial and production keys until rate limits are reached | Production use through Model Vault or Cohere deployments; no per-token price listed | Command A+ model page |
| Command A Translate tokens | Free on trial and production keys until rate limits are reached | Production use: contact [email protected] | Command A Translate model page |
| Command A Reasoning tokens | Free on trial and production keys until rate limits are reached | Production use through Model Vault | Command A Reasoning model page |
| Command A Vision tokens | Free on trial and production keys until rate limits are reached | Production use: contact [email protected] | Command A Vision model page |
| Command A tokens | $2.50 input; $10 output | Per 1M tokens, as listed on the model page; not shown on the public pricing page | Command A model page |
| Model Vault H100 80GB | $4.25 per GPU hour (promotional; on-demand $6.25) | Promotional prices stated as valid through 31 March 2027 | Pricing page |
| Model Vault B200 192GB | $8.00 per GPU hour (promotional; on-demand $9.75) | Pricing is based on the GPUs a model runs on and the number of replicas | Pricing page |
| Trial keys | Free | Rate limited and not permitted for production or commercial purposes | Pricing page |
Cohere's public pricing page lists per-token rates only for legacy models offered to existing customers (Command, Command Light, Command R and R+ versions and Aya Expanse). Cached-input and batch prices are not stated by Cohere on the pages checked (2026-11-15). For a cross-provider view, see the LLM Pricing Tracker and Model Pricing Explained.
Context window and output limits
As of 2026-11-15. Source: Cohere models page and model pages.
| Limit | Value | Notes | Source |
|---|---|---|---|
| Command A+ context / output | 128K / 64K tokens | Largest output limit in the family | Models page; Command A+ model page |
| Command A Translate context / output | 8K / 8K tokens | The page description says 16K total (8K input and 8K output); the specifications list 8,000 | Command A Translate model page |
| Command A Reasoning context / output | 256K / 32K tokens | Thinking can be enabled or disabled, with token budgets in Cohere's Reasoning guide | Command A Reasoning model page |
| Command A Vision context / output | 128K / 8K tokens | Up to 20 images per request | Command A Vision model page |
| Command A context / output | 256K / 8K tokens | Hugging Face config defaults to 128K for self-hosting | Command A model page; model card |
For RAG, the Command A card recommends document snippets of about 100 to 400 words.
Modalities and tools
Every model page lists the same capability set: reasoning, multilingual, safety modes, citations, tool use, structured outputs and image inputs. The models page lists image input only for Command A+ and Command A Vision. Page-level details:
- Tool use and agents: Command A works with search engines, APIs and vector databases and supports multistep tool use for ReAct-style agents; its page says it is good at avoiding unnecessary tool calls. Command A+ has conversational tool-use training and, per its page, can "autonomously take actions and interact with its environment".
- Vision: Command A Vision accepts up to 20 images per request and does not generate images. Its Limitations section says tool use is not supported, which differs from the capability list on the same page.
- Reasoning: Command A Reasoning and Command A+ allow thinking to be turned on or off, with token budgets documented in the Reasoning guide.
- Safety modes: Contextual mode (default) rejects harmful or illegal suggestions while allowing wide-ranging interactions; Strict mode avoids sensitive topics such as violence, sexual content and profanity (Command A card).
Where is Cohere Command A available?
As of 2026-11-15. Source: Cohere models page, model pages, changelog and data commitments page.
| Channel | Details | Source |
|---|---|---|
| Cohere API and Playground | All five models; Chat V2, Chat V1 and Chat Completions endpoints | Model pages |
| Model Vault | Production use of Command A+ and Command A Reasoning on dedicated GPU capacity | Command A+ and Reasoning model pages |
| Private deployment | On-premises or isolated environments; offered for Command A+ to enterprise customers and for Command A Translate for sensitive data | Changelog; Translate model page; data commitments |
| Amazon Bedrock and SageMaker | Command A: Bedrock listed as "Coming Soon"; SageMaker ID unique per deployment | Models page |
| Microsoft Azure AI Foundry | Command A (ID unique per deployment); Command A+ as coherelabs-command-a-plus-05-2026-w4a4 | Models page |
| Oracle OCI | Command A as cohere.command-a-03-2025 | Models page |
| OpenRouter | cohere/command-a and cohere/command-a-plus | Models page |
| Hugging Face (self-host) | CohereLabs/c4ai-command-a-03-2025; CohereLabs/command-a-plus-05-2026 in BF16, FP8 and W4A4 builds | Cohere Labs model cards |
The models page lists no third-party platform IDs for Command A Reasoning, Command A Vision or Command A Translate. Regional availability is not stated by Cohere on the pages checked (2026-11-15). For a side-by-side of cloud-hosted enterprise models, see Amazon Nova vs Microsoft MAI vs Cohere Command.
Hardware for self-hosting
| Model | Stated hardware | Source |
|---|---|---|
| Command A+ | BF16: 4x B200 or 8x H100; FP8: 2x B200 or 4x H100; W4A4: 1x B200 or 2x H100 (W4A4 recommended for most uses) | Command A+ model card |
| Command A Translate | One or two A100 or H100 GPUs | Translate model page |
| Command A Reasoning | 4x H100 for production; 4x A100 for tryouts and evaluations | Reasoning model page |
| Command A Vision | Not stated by Cohere on the pages checked (2026-11-15) | Vision model page |
| Command A | Two A100 or H100 GPUs | Command A model page |
Access restrictions
Trial keys are free but rate limited and may not be used for production or commercial purposes. Production keys are needed for commercial work, and for four of the five models Cohere directs production use to Model Vault or its sales team. Cohere deployments are governed by commercial agreements and Cohere's Usage Policy.
Self-hosting depends on the licence. Command A+ weights are under Apache 2.0, which permits commercial use. Command A weights are under CC-BY-NC 4.0, a non-commercial licence, and use also requires adherence to the Cohere Labs Acceptable Use Policy, so commercial use of Command A runs through Cohere's API or partner clouds. See the Open-Weight Model Licence Tracker.
Benchmarks (provider-reported, dated)
As of 2026-11-15. Provider-reported figures only, as shown on the pages checked. Independent leaderboard scores (LMArena, Artificial Analysis, Epoch AI) were not checked for this profile and are not shown.
| Model | Figure | Value | Source (checked 2026-11-15) |
|---|---|---|---|
| Command A+ | Throughput vs Command A Reasoning | Up to 110% higher | Changelog, 20 May 2026 |
| Command A+, Translate, Reasoning, Vision | Benchmark scores | Not stated by Cohere on the pages checked | Model pages; Command A+ model card |
| Command A | GPQA Diamond | 50.51 | Command A model card |
| Command A | Inference throughput vs Command R+ 08-2024 | 150% higher | Command A model page |
The Command A+ page describes it as the "fastest and most performant model in the Command A family by far" without supporting figures on that page.
Model card and system card
Cohere Labs publishes Hugging Face model cards for Command A and Command A+. The Command A card covers tool use and RAG chat templates, citations and the two safety modes. The Command A+ card describes a sparse Mixture of Experts decoder-only Transformer with 128 experts, 8 active per token and one shared expert, with sliding-window and global attention layers interleaved at a 3:1 ratio, plus the three quantised builds. Neither card lists safety evaluation results. A separate system card is not stated by Cohere on the pages checked (2026-11-15); the Command A model page links the Command A Technical Report from its Responsible Use section.
Data handling
Cohere's enterprise data commitments page sets the terms by deployment type:
- SaaS platform retention: logged prompts and generations are deleted automatically after 30 days, with exceptions for legal or contractual requirements and usage flagged as potentially violating Cohere's terms or Usage Policy.
- Training use: enterprise customers can opt out in the dashboard under Settings > Data Controls; the change applies to data created afterwards. If opted in, personal information is filtered out before data is used for training.
- Zero data retention: available to enterprise customers with additional usage commitments; Cohere then logs no prompts or generations, though usage data is still collected.
- Private and partner cloud deployments: Cohere does not receive customer prompts or generations.
- Assurance: the Trust Center provides ISO 27001 and ISO 42001 certificates and the SOC 2 Type II report; a Data Processing Addendum is available for SaaS customers.
Treat Cohere API keys as production secrets. See AI model provider API keys and, for self-hosted weights, the Self-Hosted LLM Security Guide.
Deprecation status
As of 2026-11-15. Source: Cohere models page and changelog.
| Model or item | Details | Source |
|---|---|---|
| command-a-plus-05-2026 | Live; no deprecation notice | Models page |
| command-a-translate-08-2025 | Live; no deprecation notice | Models page |
| command-a-reasoning-08-2025 | Live; no deprecation notice; recommended as a chat replacement in the 4 April 2026 retirement notice | Models page; changelog |
| command-a-vision-07-2025 | Live; no deprecation notice | Models page |
| command-a-03-2025 | Live; no deprecation notice; recommended as a chat replacement in the 4 April 2026 retirement notice | Models page; changelog |
| Predecessors (command, command-light, Command R 03-2024, Command R+ 04-2024, command-r and command-r-plus aliases) | Deprecated 15 September 2025 | Models page |
Because Cohere calls Command A+ the last model in the family, future Command releases will carry new names.
How Cohere Command A differs from other models in its family
| Model | Architecture | Context / output | Input | Positioning (provider wording) |
|---|---|---|---|---|
| Command A+ | MoE, 218B total, 25B active | 128K / 64K | Text, image | Vision, agentic, reasoning and translation in one set of weights |
| Command A Translate | 111B | 8K / 8K | Text | Machine translation across 23 languages |
| Command A Reasoning | 111B | 256K / 32K | Text | Thinks before answering, for agentic and nuanced problem solving |
| Command A Vision | Not stated by Cohere | 128K / 8K | Text, image | Charts, diagrams, tables, OCR and document question answering |
| Command A | 111B dense | 256K / 8K | Text | Tool use, agents, RAG and multilingual work |
The documented differences are architecture (only A+ is a Mixture of Experts), licence (Apache 2.0 for A+ and CC-BY-NC 4.0 for A), knowledge cutoff (April 2025 for A+, June 2024 for the rest), language coverage (48 for A+) and output limit. Workloads that need the full 256K context fit Command A or Command A Reasoning; Command A+ caps input at 128K. For other providers' current models, see Frontier AI Models Compared and Open-Weight Frontier Models Compared.
Practitioner checklist
- Pin the full dated model ID (for example command-a-plus-05-2026) in code and record the date you adopted it.
- Check the licence before self-hosting: Apache 2.0 for Command A+, CC-BY-NC 4.0 plus the Acceptable Use Policy for Command A.
- Confirm production pricing in writing with Cohere for every model other than Command A, and keep trial keys out of production.
- Size inputs to the model: 128K for A+ and Vision, 256K for A and Reasoning, 8K for Translate, and chunk documents accordingly.
- Test tool calling on Command A Vision before relying on it, since its page lists tool use as both supported and unsupported.
- Decide on training opt-out and zero data retention before sending regulated data to the SaaS platform; prefer private deployment where Cohere must not receive prompts.
- Store Cohere API keys in a secrets manager, separate keys by environment and rotate them on staff changes.
- Watch the Cohere changelog and deprecations page monthly for changes to the five Command A IDs.
Standards and references
- Cohere: Models overview (accessed 2026-11-15)
- Cohere: Command A+ model page (accessed 2026-11-15)
- Cohere: Command A model page (accessed 2026-11-15)
- Cohere: Command A Reasoning model page (accessed 2026-11-15)
- Cohere: Command A Vision model page (accessed 2026-11-15)
- Cohere: Command A Translate model page (accessed 2026-11-15)
- Cohere Labs: Command A+ model card (accessed 2026-11-15)
- Cohere Labs: Command A model card (accessed 2026-11-15)
- Cohere: Changelog (accessed 2026-11-15)
- Cohere: Pricing (accessed 2026-11-15)
- Cohere: Enterprise data commitments (accessed 2026-11-15)
Related NHI Mgmt Group resources: Cohere Models, Amazon Nova vs Microsoft MAI vs Cohere Command, Open-Weight Model Licence Tracker, Self-Hosted LLM Security Guide, AI model provider API keys.