Cohere AI pricing in 2026: A complete guide to real costs

A first look at the pricing page

Before we get into numbers, here's cohere.com/pricing itself. Three tabs (Workplace systems, Generative models, Advanced retrieval models), a Model Vault table, an FAQ block with legacy rates, and a lead-capture form.

If you've been on the page recently, you'll have noticed something odd: the per-token tables for current generative and retrieval models don't show up on a "view source" of the HTML. They're rendered client-side via Sanity CMS, which means scraping tools (and most LLM crawlers) get only the static parts: the Model Vault dedicated-instance table, the legacy/Aya callouts in the FAQ, and the enterprise "talk to sales" cards. For a public pricing page from a vendor whose pitch is transparency, it's a strange choice. We've cross-referenced OpenRouter, AWS Bedrock, Cohere's own docs and the Wayback Machine snapshot to get the rest, and that's what this post is built on.

Here's the price snapshot at a glance:

Cohere API pricing, model by model

Cohere's catalogue splits into three buckets: generative (the Command family), retrieval (Embed and Rerank), and audio (Transcribe). Most of these have a per-token or per-search rate; a few don't.

Generative: the Command family

These are Cohere's text-generation models, served through the Chat endpoint. The full spec (status, modality, context window, output cap) comes from the Cohere models catalogue. Per-token rates for the legacy block are verbatim from the Cohere pricing FAQ; for the August 2024 R / R+ refresh and Command A, we used OpenRouter (which passes Cohere's first-party prices through directly).

A few things worth noticing. First, Command R+'s effective price has come down: the 04-2024 launch listed at $3.00 / $15.00, the 08-2024 refresh is at $2.50 / $10.00, and that's where it sits today. The original Hacker News launch thread anchored R+ at the higher rate and the community pushed back on whether the parameter jump justified it; the price cut a few months later is the answer.

Second, Command A is roughly a "modern R+ on the same rate card". Same $2.50 / $10.00, but a 256k context, 150% higher throughput, and better agentic performance per the docs. If you were on R+ for general work, A is the new default.

Third, Command R7B is genuinely cheap. At $0.0375 in / $0.15 out per 1M, you can run an awful lot of summarisation, classification, or routing for the price of a single GPT-class request.

Retrieval: Embed and Rerank

This is where Cohere's pricing really earns its reputation. The retrieval stack is what most of the developer praise on PeerSpot and dev blogs is about, and the rates back that up.

Two definitions worth nailing down here:

• A "search" is one query plus up to 100 documents, per the Cohere pricing FAQ. Anything over 500 tokens gets auto-chunked into multiple documents and each chunk counts.
• Embed v4 handles PDFs natively. That's actually a meaningful pricing detail, because the alternative on other vendors is parsing PDFs into chunks yourself and embedding each chunk separately.

What we'd reach for: Embed v4 + Rerank v3.5 + Command R is the canonical "cheap, capable RAG stack" on Cohere, and the end-to-end cost on a moderate retrieval workload comes out under what a comparable GPT-class + OpenAI embeddings pipeline would cost. If you go heavier on quality, swap to Rerank 4 Pro and Command A.

Audio: Cohere Transcribe

Transcribe is Cohere's speech-to-text model: cohere-transcribe-03-2026, 14 languages, 25 MB max file. Per-minute pricing is on the live pricing tab and isn't exposed in the static HTML, so we can't quote it here without misleading you. The audio transcription docs confirm trial keys are capped at 5 requests per minute and production access is via sales@cohere.com.

What's not on the public pricing page

This is the bit that's worth being honest about, because it's the difference between Cohere and a vendor like Anthropic or OpenAI where every model has a published per-token rate.

The following are listed on Cohere's live pricing page but don't render in static HTML and aren't surfaced cleanly by third-party hosts either:

• Command A+, Cohere's flagship MoE model, released open-source on 19 May 2026. Marketed for "sovereign critical infrastructure". No public per-token rate.
• Command A Reasoning, Translate, and Vision, all "contact sales" on production keys per the rate-limits docs. Trial keys cap at 20 req/min and 1,000 calls/month.
• Fine-tuning fees (training, hosting, serving).
• Per-image pricing for Command A Vision and Command A+ image inputs.
• Batch API discounts, listed but no explicit rate.
• North-Mini-Code-1.0, appears in rate-limit tables; per-token rate not exposed.

What this means for a buyer: the moment you want to compare Cohere's current flagship against GPT-5.x or Claude Sonnet on a $/1M-token basis, you can't, without booking a sales call. Whether that's a deal-breaker depends on your role. For an enterprise architect drafting a procurement deck, it's a friction point but not a wall. For a small team evaluating five LLM providers in one afternoon, it's enough to bump Cohere down the list.

Model Vault: dedicated deployments, exact prices

The Model Vault is Cohere's dedicated-instance offering: you reserve compute that runs only your models, no shared tenancy. This is also the only block of Cohere's pricing page where the actual dollar numbers are in the HTML, so we can quote it directly.

A few things to flag:

• Billing can be hourly or commit-based (monthly or annual). The monthly rate works out to ~$8.33/hour at 30×24, so the committed monthly rate is cheaper than hourly only if your usage clears about 60 to 80% of the month.
• A separate Sanity CMS data point lists a Compass-tied Model Vault price of $3.75/hour per instance, which doesn't appear in the main table above. It's worth raising in a sales conversation if Compass is in your plan.
• No on-demand Command tier in the Model Vault. If you want dedicated generative throughput, you're in custom deployment territory.

Model Vault is the answer to a specific buyer: "I need predictable monthly spend on retrieval at scale, and I don't want my workloads sharing a multi-tenant pool." A typical RAG team running, say, 500 embed-jobs/hour and 5,000 rerank queries/minute would lean towards a Rerank 4 Pro Medium ($3,250/mo) plus an Embed 4 Medium ($3,250/mo), about $6,500/month total, paid as a flat line item instead of metered per-request.

North and Compass: the enterprise platforms

The first tab of Cohere's pricing page is Workplace systems, and it's the most opaque part of the whole pricing story. Two products live there:

• North, Cohere's agent platform. Pitched on the homepage as "Your sovereign AI workplace". Connects to your tools, lets users run automations, and includes intelligent search.
• Compass, Cohere's enterprise search and discovery system, with pre-built data connectors, document parsing, and a managed index.

Neither has a public price. Both are "Get in touch for custom enterprise pricing" with a "Request a demo" CTA. The lead-capture form on cohere.com/pricing exposes a deployment-preference dropdown (AWS, Azure, GCP, OCI, IBM Cloud, Tencent Cloud, Alibaba Cloud, Private Deployment, Cohere Infrastructure (SaaS), Other), which is a fair signal of how custom these deals tend to be.

What we'd expect, based on conversations with enterprise buyers and public reference deals (Fujitsu, Oracle, RBC, Dell, LG CNS, all on the Cohere homepage logo strip): a multi-stakeholder procurement cycle, a six-figure minimum on a year-one commit, dedicated solutions-engineering support, and a custom deployment surface. If you're a Fortune 500 with a sovereign-data mandate, that's table stakes. If you're a 50-person support team that just wants an AI that triages tickets, this isn't your fit.

Reviewers on PeerSpot flag the same thing: "Cohere offers enterprise pricing for high-volume customers, and you should contact their sales team for custom pricing if you're processing billions of tokens monthly or need dedicated support, SLAs, or private deployments."

Pricing on partner clouds

Cohere is sold through every major cloud marketplace: AWS Bedrock, Amazon SageMaker, Microsoft Azure (AI Foundry), Oracle Cloud Infrastructure, Google Cloud, IBM Cloud, Tencent Cloud, Alibaba Cloud. Pricing on each is set by the marketplace, not Cohere, and isn't always identical to cohere.com.

The most surprising thing is that AWS Bedrock has been pared back. The current on-demand Cohere catalogue is just Embed 4 and Rerank 3.5; Command R, R+, and A aren't on the on-demand list anymore. The Command family on Bedrock is now Provisioned Throughput only, which is a very different cost shape:

That $49.50/hour per model unit (no commit) for Cohere Command works out to ~$29,462/month per unit. PeerSpot's "expensive to use all Oracle services" quote is the same story in another tab: cloud-marketplace markup is where the budgeting surprises live.

Azure AI Foundry sells Cohere-command-a, Cohere-rerank-v4.0-pro/fast, and embed-v-4-0 directly, but per-token rates aren't on the docs page. Oracle OCI's pretrained-models doc lists every Cohere model OCI hosts (Command A Reasoning, A Vision, A, R+, R, Embed v4 + v3 variants, Rerank 4, Rerank 3.5), but again the rates are on a separate pricing page.

The rule of thumb: if you're cost-sensitive, run Cohere on Cohere first-party. The marketplaces are convenient if you already live there, but the markup is real.

Trial vs production: the small print

A few mechanics worth knowing before you sign up, all from the official rate-limits docs and the pricing FAQ:

• Trial keys are free, but capped at 1,000 API calls per month and 20 req/min per chat model. They're explicitly not allowed for production or commercial use.
• Production keys are pay-as-you-go, with higher rate limits (500 req/min for Command A, R, R+, R7B).
• For newer model variants (Command A+, A Reasoning, A Translate, A Vision), production keys behave like trial keys. You have to contact sales@cohere.com for real production access. Easy to miss.
• Billing cadence: end of every calendar month or when your outstanding balance reaches $250, whichever comes first.
• You're only charged for billed_units, not the underlying token count. Cohere absorbs the cost of any control tokens added under the hood. Their docs include an example where the actual tokens count is 7,596 / 645 but the billed_units is 6,772 / 248. Small but real saving.
• Accounts start as personal; you need to set up an organisation if you want shared billing across teammates.

The trial-vs-production gating on flagship models is the friction point. Most LLM vendors let you hit GPT-5 or Claude Sonnet on a self-serve key from day one. Cohere doesn't, and if you're benchmarking models against each other, that asymmetry can quietly steer your testing towards whichever model you can actually run.

Worked cost examples

Let's get concrete. Three scenarios:

Scenario 1: Small RAG pipeline (cheap and cheerful)

You're a 10-person ops team, building an internal "ask the wiki" bot. ~5M tokens/day of indexing, ~50,000 retrieval queries/day, generation done with Command R.

Rerank is the dominant line. If you can collapse to top-5 rerank instead of top-100 per query, that bill comes down sharply.

Scenario 2: Mid-size support automation (the typical pitch)

You're a support team, ~1,000 tickets/day, building a deflection bot on Cohere directly. Average ticket = 3,000 input tokens (history + KB context) + 500 output tokens.

The $575 in tokens is the smallest line on the bill. The actual cost is the engineer writing the prompts, the orchestration, the retries, the eval harness, and the on-call rotation when the bot escalates wrong. Which leads to…

Scenario 3: Enterprise dedicated (Model Vault)

You're an enterprise running RAG over a 5M-document corpus, with predictable retrieval volume.

Add a North or Compass contract on top and you're in the multi-six-figure range over a year. That's what the Fujitsu and Oracle reference deals point at.

What users actually say about the bill

We pulled real practitioner voices from PeerSpot and the Hacker News Command R+ launch thread. The split is consistent and worth knowing:

"Cohere has a free tier… you can use the API in development mode. But if you go to production, you will have to pay… it can be expensive." CollinsOmondi, Mobile Developer, PeerSpot pricing thread

"Compared to models available in the market, Cohere's pricing, setup cost, and licensing are better." Shivam Singh, Senior Solution Architect, PeerSpot pricing thread

"I've used Cohere's Embed English v3.0 for RAG tasks and found it faster, cheaper, and more responsive than alternatives." PeerSpot reviewer summary

"My experience with pricing, setup cost, and licensing is that it is expensive to use all Oracle services." Senior Data Scientist, PeerSpot pricing thread

And from the Hacker News R+ launch thread (which had a much sharper edge):

"[Command-R] ended up lobotomized when the agent relied on its results." irthomasthomas, Hacker News thread

What we'd pull from this:

• The Embed and Rerank story is uniformly positive. "Faster, cheaper, more responsive" comes up over and over in retrieval contexts. This is Cohere's strongest pricing story, and it's not really being told loudly enough on the marketing page.
• Command R is where the price-performance fans live. $0.15 / $0.60 for a 35B-parameter, 128k-context model is genuinely good.
• Command R+ catches flak. The HN thread anchored R+ at a 6x price step over R, and the community wanted to know why. Simon Willison's reply settled on "it's a 104B model, you're paying for parameter count, not a new capability tier", which is honest, but not a great answer for buyers who care about output quality first.
• The real "sticker shock" isn't the per-token rate, it's the TCO. Multiple practitioner write-ups make the same point: "You need a team of developers to build the app, data scientists to tweak the models, and an IT team to manage everything."

That last one is the bit worth sitting with.

Where Cohere fits, where it doesn't

Pricing isn't really the question. Fit is. Cohere's pricing is designed for a specific buyer profile, and if you're not that buyer, the question of whether the per-token rate is "good" or "expensive" misses the point.

Cohere is a strong pick if you're an enterprise with:

• A data-sovereignty mandate (financial services, public sector, healthcare, defence, telco).
• An existing engineering team that can integrate at the API level: write the prompts, build the orchestration, run the evals.
• A primary workload that's RAG or retrieval-heavy. The Embed + Rerank + Command R stack is genuinely a price-performance win, especially on multilingual content (49 languages on Command, see the homepage).
• A multi-cloud or on-premises requirement. Cohere's deployment story (VPC, on-prem, Model Vault, partner clouds) is built for this.

Cohere is the wrong pick if you're:

• A small team that wants something working on day one, not in three months after the integration build.
• A support, ops, or content team where the unit of value is "ticket handled" or "post published", not "token processed".
• Cost-sensitive in a self-serve way. The flagship rates being gated behind a sales call means you can't actually run the comparison you want to run.
• Looking for plug-and-play AI agents inside helpdesks like Zendesk, Freshdesk, or Gorgias.

For that second profile, which is most of what we see in real buyer conversations, the question isn't "is Cohere cheaper than OpenAI per token?". It's "can I get the AI value without paying for the integration team?". And that's a different shape of pricing.

Try eesel

If you got this far and you're a support, ops, or content team weighing Cohere against a turnkey alternative, here's the honest pitch.

eesel gives you AI agents that live inside the tools you already use: Zendesk, Freshdesk, Gorgias, Slack, Gmail, Shopify, and a hundred others. No prompts to engineer, no orchestration to build, no production-key sales call. You connect your helpdesk, brief the agent in plain language the way you'd brief a new hire, and it starts handling tickets.

The pricing is also a different shape: $0.40 per ticket as a regular task, no platform fee, no per-seat fees, no monthly minimum. A 1,000-ticket/month team pays $400. A 100-ticket/month team pays $40. You get a $50 free credit (plus 2 free blog generations) before any card is required. Full pricing here.

We're not the same product as Cohere: Cohere sells models; eesel sells AI teammates that run on top of models. If your job is to ship infrastructure for a Fortune 500's sovereign AI stack, Cohere's the right call. If your job is to deflect tier-1 tickets next week, try eesel: it's a few minutes to onboard and the per-ticket math is easier to defend in a budget meeting than a Provisioned Throughput line.