Cohere Trial Key Limits: What the Free Key Allows, and What It Forbids (2026)

Most providers make free access a tier of the same account. Cohere makes it a different credential, and that design choice is the single most useful thing to understand about it. A trial key and a production key are separate objects with separate limits and separate terms — you do not upgrade a trial key, you replace it.

That is clarifying in a way other providers' free tiers are not. There is no ambiguity about which side of the line you are on, and no gradual slide from evaluation into production without noticing. But it also creates the specific failure this guide exists to prevent: teams build, test and ship on a trial key, then discover at launch that the credential in their environment variables was never licensed for what they are now doing with it.

The second Cohere-specific point is that the product surface is wider than a chat endpoint. Generation, embeddings and rerank are metered differently — rerank in search units rather than tokens — so 'am I within the trial limits?' is three questions here, not one, and the answers do not move together.

Trial Keys and Production Keys Are Different Objects

The distinction is enforced at the credential level. A trial key is rate-limited well below production allowances by design, and it is not intended — or licensed — for commercial use. A production key is a separate credential with its own limits and its own terms, created deliberately once you have a payment method attached.

The practical consequence is that the migration is a swap rather than a setting, and the swap has to reach every place the credential lives: local environments, CI secrets, staging, the deployment platform, and any teammate's machine that still has the old value in a dotfile. A partially completed swap is the classic cause of the mystifying incident where production is fine and one background worker is being throttled into the ground.

Instrument this rather than trusting the rollout. Log which key class served each request, and alert on any production traffic served by a trial credential. That single check has caught more launch-day incidents than any amount of documentation about which key to use.

The Commercial-Use Restriction Is the Real Limit

Rate limits are an engineering problem; a licence restriction is not, and this is where Cohere differs most sharply from providers whose free tier is merely smaller. Trial access is for evaluation and development. Using it to serve a product to customers is outside the terms regardless of whether your volume happens to fit inside the allowance.

Teams get here honestly. A prototype works, a demo gets shown, the demo becomes the thing, and nobody re-reads the terms attached to the key that has been in the environment file since the first week. The volume never triggers a hard stop, so there is no technical moment that forces the question — which is exactly why it needs to be a checklist item at launch rather than something you notice later.

Make it explicit: no external user touches a response generated by a trial key. That rule is easy to state, easy to audit if you have the key-class logging above, and it removes an entire category of problem that has nothing to do with capacity.

Three Endpoints, Three Meters

Cohere's surface is broader than chat, and the trial limits apply per endpoint family rather than as one pooled allowance. Generation is metered in tokens the way you would expect. Embeddings are metered in tokens too, but the usage pattern is entirely different — a single corpus indexing job can consume more in an hour than weeks of interactive chat traffic. Rerank is metered in search units, which do not convert to tokens at all and therefore do not appear in any token-counting instrumentation you may have brought from another provider.

The consequence is that a trial evaluation of a retrieval pipeline hits its ceiling in a place teams do not expect. The chat allowance is fine, the embeddings allowance is exhausted by the initial index build, and the rerank allowance is being consumed per query in a unit nobody is tracking. Track all three separately from the start.

There is an upside specific to embeddings: for a fixed model and input, embeddings are deterministic, which makes them genuinely cacheable in a way completions are not. Persisting embeddings rather than recomputing them on every run is the single highest-leverage thing you can do to make a trial allowance last through a real evaluation.

Trial Limit or Actual Outage?

The usual ambiguity applies — rising errors and rising latency look the same whatever the cause — with one Cohere-specific twist that makes diagnosis easier if you use it.

Read the error body first: a trial-key rejection names the limit it enforced and typically carries a reset hint, while an incident produces generic upstream errors, timeouts and 5xx responses with nothing account-specific in them. Then correlate with your own traffic, since your limit tracks your peaks and an outage does not. The Cohere-specific check is the cross-endpoint one: if generation is failing while embeddings on the same key succeed, that is per-endpoint trial limiting rather than a platform incident, because an incident on api.cohere.com is unlikely to respect your endpoint boundaries so neatly.

On a trial key there is no support path you can escalate down inside an incident window, which makes an independent probe the only source of truth you actually control. It runs on its own schedule and is unaffected by your traffic, so it answers the first question of any incident — is this us or is this them — without you having to reason about it.

When to Get a Production Key

Earlier than most teams do, and the trigger is not volume. Get one the moment anyone outside your team can see a response, because that is the moment the licence question becomes real regardless of how much allowance is left. Get one before a launch date, not on it, so a partially propagated credential swap is discovered in staging rather than in production. And get one before you benchmark for capacity planning, because trial-key throughput tells you nothing reliable about production behaviour — the limits are different by design, so any numbers you gathered need re-measuring on the credential you will actually ship.

The swap itself is mechanical: create the production key, roll it through every environment, verify by logging key class per request, then revoke the trial key so it cannot silently serve traffic again. That last step is the one people skip, and it is the reason trial credentials turn up in production months later.

After the swap, keep independent monitoring on api.cohere.com. Moving to production limits removes the ceiling you were hitting; it does not remove incidents, and once real users depend on generation, embeddings or rerank, telling your problem from theirs quickly is worth more than it was during evaluation.

Frequently Asked Questions

Is Cohere free to use?

There is free access, but it is deliberately narrow: a trial key, which is a separate credential from a production key, rate-limited well below production allowances by design and intended for evaluation and development rather than commercial use. That design is actually clearer than a graduated free tier, because there is no gradual slide from prototype into production without noticing — you cannot upgrade a trial key, you replace it. The risk it creates is the mirror image: teams ship with the credential they started with and only later discover it was never licensed for what they are doing.

What is the difference between a Cohere trial key and a production key?

They are different objects, not different settings on the same one. A trial key carries lower rate limits and terms restricted to evaluation and development; a production key is created separately once a payment method is attached, and carries its own limits and terms. Because it is a swap rather than an upgrade, the migration has to reach every place the credential lives — CI secrets, staging, the deploy platform, developers' local environments — and a partially completed swap is the classic cause of one background worker being throttled while everything else looks healthy.

Can I use a Cohere trial key in production?

No. The restriction is about licence rather than capacity, so it applies regardless of whether your volume happens to fit inside the trial allowance — and because there is no technical hard stop at the moment you cross the line, nothing forces the question. Make it a launch checklist item with a simple rule: no external user sees a response generated by a trial key. If you log which key class served each request, that rule is auditable in one query rather than being a matter of memory.

What are the Cohere trial key rate limits?

They are published per endpoint and change over time, so read them from the current documentation for the endpoints you are actually calling. The more durable point is that there are three meters, not one: generation and embeddings are metered in tokens, while rerank is metered in search units that do not convert to tokens at all and therefore never appear in token-counting instrumentation. A retrieval evaluation typically exhausts embeddings during the initial index build while the chat allowance is barely touched.

Why did my Cohere embeddings quota run out so quickly?

Because indexing a corpus is a fundamentally different usage shape from interactive traffic — a single index build can consume more in an hour than weeks of chat. The good news is specific to embeddings: for a fixed model and input they are deterministic, which makes them genuinely cacheable in a way completions are not. Persisting embeddings to storage and recomputing only what changed is the highest-leverage change available for making a trial allowance survive a real evaluation, and it remains the right architecture after you move to a production key.

How do I tell a Cohere trial limit from an outage?

Read the error body: a trial-key rejection names the limit and usually carries a reset hint, while an incident produces generic upstream errors, timeouts and 5xx responses with nothing specific to your account. Correlate with your own traffic, since your limit tracks your peaks and an outage does not. Then use the Cohere-specific check — try a different endpoint family on the same key. Generation failing while embeddings succeed points at per-endpoint trial limiting, because a genuine incident on api.cohere.com is unlikely to respect your endpoint boundaries that precisely.

When should I move from a Cohere trial key to production?

The trigger is exposure, not volume. Move the moment anyone outside your team can see a response, because that is when the licence question becomes real whatever the allowance says. Move before a launch date rather than on it, so an incomplete credential rollout surfaces in staging. And move before any capacity benchmarking, since trial throughput is different by design and tells you nothing reliable about production. Finish by revoking the trial key — the step teams skip, and the reason trial credentials resurface in production months later.

Related Guides

Trial Key Limit, or Is Cohere Actually Down?

A trial key gives you no support path to escalate down inside an incident window. API Status Check probes api.cohere.com independently of your traffic and alerts on errors and latency, so you can separate your ceiling from their outage in seconds.

Start Your Free Trial →