Cohere’s API is the self-serve way to use its models. Cohere distinguishes two kinds of API key: evaluation keys that are free but limited, and production keys that are paid and much less limited.[3] Every account receives a Trial key automatically when it is created, and a Production key is obtained through a Go to Production workflow.[1] The distinction matters to licensing because it is a use right as well as a price: trial keys are not permitted to be used for production or commercial purposes.[1]
Editions: Trial and Production keys
Trial key. API calls made with a Trial key are free, rate limited, and not permitted for production or commercial use.[1] Trial keys, and production keys used with newer Chat model variants, are limited to 1,000 API calls a month.[3] Data sent on a Trial key is outside the Enterprise Data Commitments: those cover paying customers, which for the SaaS platform means a credit card on file, and trial usage falls under the Terms of Use and Privacy Policy.[14]
Production key. A Production key is billed pay-as-you-go.[1] Only a user with Owner privileges can request one, from the Billing and Usage page.[1] Every account begins as a personal account with access to Trial keys only; an organization is a group of personal accounts sharing one billing portal, and organizations are not automatically given production access.[1]
Going live
The Go to Production workflow requires acknowledging Cohere’s SaaS agreement and terms of service, and recognising model limitations, model cards and the data statement.[4] The applicant is asked whether the usage involves any sensitive use case in Cohere’s usage guidelines. If it does, the production key may be rate limited like a trial key until Cohere’s safety team manually approves the use case, a review that Cohere says takes no longer than 72 business hours.[4] The catalog row is Go to Production workflow.
Metrics: tokens, searches and pages
Cohere’s generative models, including Command A+, Command A, Command R7B, Command R and Command R+, are priced per token, with input tokens and output tokens priced differently.[2] Rerank models are priced by the quantity of searches and Embed models by the number of tokens embedded.[2]
Billed tokens
The Chat response reports a count of billed tokens alongside the generic token count. Billed tokens are what is charged. They can be lower than the generic count because Cohere sometimes adds tokens under the hood and a model may emit special tokens, and the customer is not charged for tokens it cannot control.[2] For reconciliation against invoices, use billed_units, not the total token fields. The metric is Chat tokens and its embedding counterpart is Embed tokens.
Rerank search units
Cohere defines a single search unit as one query with up to 100 documents to be ranked. If any document exceeds 500 tokens, including the length of the search query, it is split into chunks, and each chunk counts as an individual document toward the 100.[1] A query over many long documents can therefore consume more than one search unit. The metric is Rerank search unit.
Published list prices
The model pages show these per-token prices at the retrieval date.
| Model | Input per 1M tokens | Output per 1M tokens | Source |
|---|---|---|---|
| Command A (command-a-03-2025) | USD 2.50 | USD 10 | Model card[5] |
| Command R (08-2024) | USD 0.15 | USD 0.60 | Model card[6] |
| Command R7B | USD 0.0375 | USD 0.15 | Model card[7] |
| Command R+ (08-2024) | USD 2.50 | USD 10 | Model card[8] |
| Aya Expanse (API) | USD 0.50 | USD 1.50 | Pricing page FAQ[1] |
The pricing page also lists legacy prices for existing customers: Command at USD 1.00 and USD 2.00, Command-light at USD 0.30 and USD 0.60, Command R 03-2024 at USD 0.50 and USD 1.50, Command R+ 04-2024 at USD 3.00 and USD 15.00, and Command R+ 08-2024 at USD 2.50 and USD 10.00 per million tokens.[1] The pricing page’s model cards, which load as page data, give Embed 5 Pro at USD 0.12 and Embed 5 Fast at USD 0.08 per million tokens, Rerank 4 Pro at USD 2.50 and Rerank 4 Fast at USD 2.00 per thousand searches, and Parse 5 at USD 1.50 per thousand pages.[1] Because Cohere’s agreements allow price changes on seven days notice,[13] these figures should be recorded with their retrieval date; see Cohere agreements and data terms.
Models that are free until a rate limit
Several newer models are not billed per token. Command A+, North Mini Code, Command A Reasoning, Command A Vision, Command A Translate and North Small Translate are described as free on both trial and production keys until rate limits are reached.[9][10] Production use then moves elsewhere: Command A+ and North Mini Code through Model Vault, and Command A Vision and Command A Translate through Cohere sales.[9][10] The rate-limits page confirms that production keys work like trial keys for newer variants such as Command A Reasoning, and that sales must be contacted for production use.[3]
Counting, billing and floors
Billing cycle
Production usage is billed pay-as-you-go. The bill is issued at the end of every calendar month or when outstanding balances reach USD 250.[1] Under the SaaS Agreement, fees are those in Cohere’s then-current price list unless an Order Form says otherwise, Cohere gives at least seven days notice of changes, disputed charges must be raised within 45 days, and late amounts bear interest of 1.5 percent a month.[13]
Rate limits
Cohere publishes limits per model and endpoint.
| Endpoint | Trial limit | Production limit |
|---|---|---|
| Chat: Command A, Command R+, Command R, Command R7B, North Mini Code | 20 requests per minute | 500 requests per minute |
| Chat: Command A+, A Reasoning, A Translate, A Vision | 20 requests per minute | Contact sales |
| Embed | 2,000 inputs per minute | 2,000 inputs per minute |
| Embed (images) | 5 inputs per minute | 400 inputs per minute |
| EmbedJob | 5 requests per minute | 50 requests per minute |
| Rerank | 10 requests per minute | 1,000 requests per minute |
| Audio transcriptions | 5 requests per minute | Contact sales |
| Parse | 500 requests per minute | 500 requests per minute |
The figures are from Cohere’s rate-limits page.[3] Rate limits are a capacity constraint rather than a licence metric, but they set the practical ceiling for a workload before it must move to Model Vault or a negotiated agreement. The SaaS Agreement also allows Cohere to limit calls in its sole discretion unless an Order Form sets them.[13] The catalog rows are API rate limit and Trial API key calls.
Model lifecycle and deprecation
Cohere uses four lifecycle terms: Active, Legacy, Deprecated and Shutdown. A deprecated model is no longer available to new customers but remains available to existing users until retirement, and an existing user is anyone who has used the model or endpoint within 90 days of the deprecation announcement.[12] The models page lists command-r-03-2024, command-r-plus-04-2024, command-light and command as deprecated from 15 September 2025.[11]
The deprecation history also records commercial changes. Fine-tuning for command-light, command, command-r, classify and rerank was retired effective 15 September 2025, and previously fine-tuned models are no longer accessible.[12] Earlier, rerank-english-v2.0 and rerank-multilingual-v2.0 were priced at USD 1.00 per thousand searches and shut down on 30 April 2025, and several Embed v2 models and Aya Expanse 8B were scheduled for retirement on 4 April 2026.[12] An asset manager should record the shutdown date of each model in use, because a deprecated model with no shutdown date yet still has a replacement path to plan. The catalog row is Model lifecycle and deprecation policy.
Out of scope
Negotiated Order Forms with committed spend or custom rate limits, fine-tuned Custom Models, North and Compass pricing, and the pricing of cloud marketplace listings, which the cloud provider sets.