Cohere Review 2026

What is Cohere?

Cohere's newest chat model, Command A+, carries no per-token charge on the hosted API until you hit its rate limits, while the older Command A still lists $2.50 per million input tokens and $10 per million output tokens. This Cohere API pricing and review page sorts out what you actually pay, which models are current, how far the free trial key stretches, and whether Cohere's retrieval tools justify picking it over OpenAI. Figures come from cohere.com/pricing, Cohere's model pages and its rate-limit guide, read in October 2026.

Command A+, Command A and the models Cohere sells today

Cohere is a privately held Toronto company founded in 2019 by Aidan Gomez, Ivan Zhang and Nick Frosst. It has never chased consumer chatbot users. Its customers are banks, telecoms, governments and software firms, and the API is the part developers touch first.

The current line-up:

• Command A+ (command-a-plus-05-2026), released on 20 May 2026. A mixture-of-experts model with a 128K context window, up to 64K output tokens, image input and 48 languages. Its weights are published on Hugging Face under Apache 2.0.
• Command A (the March 2025 build). 111B parameters, a 256K context window, 8K output and 23 languages, Hindi among them.
• Command A Reasoning, Vision and Translate, specialist variants added in mid-2025.
• Command R and Command R+ in their August 2024 snapshots, plus the smaller Command R7B from December 2024. All three are still live with 128K context and 4K output.
• Embed v5 (Pro and Fast), Embed v4 and the older Embed v3 family for vectors, plus Rerank 4 (Pro and Fast) and Rerank 3.5 for re-ordering search hits.
• Aya models from Cohere's research lab. The Tiny Aya set covers 70 languages, and its "Fire" variant is tuned for South Asian languages.

Parse 5 (documents), Transcribe (speech-to-text) and North Mini Code round it out. Output is text or vectors only.

Cohere API pricing for Command R, and why it is no longer the default

Command R was Cohere's best-known model in 2024, so here is the honest position. The original command-r and command-r-plus aliases (the March and April 2024 snapshots) were deprecated on 15 September 2025. The August 2024 versions remain on the API, but Cohere's own Command R page now points most users to Command A. Command R Plus (08-2024) now sits in the pricing page's legacy table, which is aimed at existing customers.

Rupee figures are conversions at roughly ₹94.7 to the dollar. Cohere quotes and bills in US dollars and has no rupee price list.

A worked example makes the spread concrete. Picture a support assistant answering 20,000 questions a month, each sending 3,000 tokens of retrieved context and receiving 500 tokens back. That is 60 million input and 10 million output tokens. On Command A the bill is $150 plus $100, so $250 a month (about ₹23,700). Run the same load on Command R7B and it drops to roughly $3.75. GPT-6.1 Sol would cost $220. The spread inside Cohere's own range dwarfs the Cohere-versus-OpenAI gap.

Production accounts are invoiced monthly, or earlier once the unpaid balance hits $250. Tokens Cohere inserts behind the scenes are not charged.

Cohere free trial key limits, and why they run out fast

Trial keys live on the API Keys page of the dashboard, and calls made with them cost nothing. Cohere's rate-limit page sets these ceilings:

• 1,000 API calls per month in total.
• 20 requests a minute on each Chat model.
• 10 requests a minute on Rerank.
• 2,000 inputs a minute on Embed for text, but only 5 a minute for images.
• 5 requests a minute for audio transcription.

The bigger catch is legal rather than technical. Cohere's terms do not allow a trial key to be used for production or commercial work. A thousand calls covers a weekend of prompt testing, not a small internal tool. The key exists for evaluation; anything real is meant to move to a billed key covered by Cohere's enterprise data commitments.

Getting a production key is not instant. An account Owner has to fill in an application from the Billing and Usage page, and members of an organisation go through the same form. Once approved, Chat on Command A, R, R+ and R7B rises to 500 requests a minute and Rerank to 1,000. The newest variants (Command A+, Reasoning, Vision and Translate) keep trial-style caps, including the 1,000-call monthly ceiling, even on a production key, and Cohere asks you to contact sales before putting them into production. For Command A+ the documented production route is Model Vault.

Cohere Rerank pricing and its place in a RAG pipeline

Rerank is billed per search rather than per token. One search is one query sent with up to 100 documents. Any document longer than 500 tokens (counting the query) is split into chunks, and each chunk counts as a document, so a query against 100 long passages can quietly become two or three searches. The hosted per-search rate sits in a tabbed section of the pricing page that did not render for this review, so confirm it on cohere.com/pricing before you budget.

What is published plainly is Model Vault, Cohere's dedicated-instance option. A Medium instance of Rerank 4 Fast, Rerank 4 Pro or Rerank 3.5 costs $5 an hour or $3,250 a month (about ₹3.08 lakh). Rerank 4 Pro on a Large instance is $10 an hour or $6,500 a month.

In a RAG setup the usual pattern runs in four steps. Pull the top 50 to 100 candidates from your vector store or keyword index, send them with the user's question to the reranker, keep the best three to five, and pass only those to Command through the documents field so the answer carries citations. Reranking trims the context sent to the generator, where most of the token bill sits. Rerank 4 accepts 32K tokens per document and handles semi-structured JSON; Rerank 3.5 stops at about 4K.

Cohere vs OpenAI embeddings: published prices against retrieval depth

OpenAI lists text-embedding-3-small at $0.02 per million tokens and text-embedding-3-large at $0.13, and those numbers are easy to find. Cohere says Embed is charged by the number of tokens embedded, but the hosted rate was not readable on its pricing page during this review. The one Embed figure that was visible is Model Vault: a Small instance of Embed 5 Pro or Embed 5 Fast costs $3 an hour or $2,000 a month (about ₹1.89 lakh).

On capability Cohere has the stronger retrieval story. Embed v4 and v5 accept up to 128K tokens per input, take mixed text and images such as PDF pages, and let you pick output sizes from 256 dimensions up to 1,536 on v4 or 2,048 on v5. The older Embed v3 models cap each input at 512 tokens, so long documents need chunking. Add Rerank and you have an enterprise search stack under one contract.

OpenAI wins on simplicity: one bill, prices in plain view, and the same SDK you already use for chat. Switching later means re-embedding your corpus, so test on your own documents first.

North 2, Model Vault and private deployment for regulated teams

Cohere's homepage now leads with North 2, the second version of its North platform for workplace agents. It bundles knowledge-searching agents, natural-language automations, memory, guardrails, connectors and an admin console. North can run in your own VPC, on your own servers or through Model Vault. Pricing is a custom quote; no figure is published.

For the models themselves there are four ways in:

• Cohere's own hosted API.
• Cloud marketplaces: Amazon Bedrock and SageMaker, Azure AI Foundry, and the OCI Generative AI service, where Command A, Command A Reasoning, Command A Vision, Embed v4 and Rerank 3.5 are listed.
• Private deployment inside your VPC.
• On-premises installs, including air-gapped sites.

For private and third-party cloud deployments, Cohere's data commitments page says it does not receive your prompts or outputs. For an Indian bank already on AWS or Oracle Cloud, that is often the easiest procurement route.

What Cohere keeps from your prompts

On the hosted platform, logged prompts and generations are deleted after 30 days, except where law, a contract or a policy violation requires longer. Enterprise customers can turn off training use under Settings, then Data Controls; the change applies only to data created afterwards. Zero data retention is offered on request to approved enterprise customers who commit to extra usage. Trial-key traffic falls under the general terms and privacy policy rather than the enterprise commitments, which is one more reason not to push real customer data through a free key.

Rough edges an Indian startup will notice

• No rupee billing. Prices and invoices are in US dollars, and Cohere lists no India office.
• Hard-to-read pricing. Hosted Embed and Rerank rates are hidden in tabs; the docs push you back to the pricing page.
• Gated production. You apply for a production key, and the newest models still need a sales conversation.
• Narrow output. Text and vectors only, with no image generation or text-to-speech.

We found no India-specific restriction or terms in Cohere's docs, so sign-up and card billing should work as they do elsewhere. Confirm GST treatment with your accountant.

A first Cohere RAG test in one afternoon

• Create an account at dashboard.cohere.com and copy the trial key from the API Keys page.
• Install an SDK: pip install cohere for Python, npm i cohere-ai for TypeScript, or the Java and Go clients.
• Prototype on command-r7b-12-2024, then compare answers on command-a-03-2025.
• Add rerank-v4.0-fast between your retriever and the chat call, counting calls against the 1,000 monthly limit.
• Before launch, apply for production access and switch off the training toggle under Data Controls.

Who should pay: the Cohere API pricing and review verdict

Cohere suits companies building enterprise search, multilingual support bots or document question-answering, especially where data cannot leave a VPC or must stay on Oracle or AWS contracts. Its embedding and reranking models are the main draw, and Command R7B is one of the cheapest capable chat models on any price list. A solo developer wanting one general model and published prices for every endpoint will be happier with OpenAI, or with Mistral, whose Small 4 costs $0.15 in and $0.60 out per million tokens.