WaveSpeedAI

Replicate Problems in 2026: Unexpected Charges, Cold Starts, Slow Support, Blocked Requests and How to Fix Them

The Replicate problems users report most: charges while idle, slow cold starts, no reply from support, blocked requests and API quirks. What causes each one and what to do about it.

By WaveSpeedAI14 min read

If your Replicate bill is higher than the predictions you ran, the usual reason is that you are paying for hardware time, not for results. Public models charge only while a request is processing. Deployments and most private models charge for every second an instance is online, including the time it spends booting and the time it sits idle. A deployment that is stuck starting up can therefore cost money without completing a single prediction.

Billing is the most common complaint, but not the only one. This guide goes through the five problems Replicate users report most often: unexpected charges, slow cold starts, support that does not reply, requests blocked by moderation, and API behavior that differs from model to model. For each one it explains the cause, what you can do on Replicate, and what to choose if you would rather not deal with it at all.

Sources checked November 22, 2026: Replicate’s billing, prepaid credit and deployment documentation, Replicate’s public GitHub repositories and its Trustpilot page.

The five problems at a glance

ProblemCauseQuick fix
Charged while nothing runsDeployments and private models bill setup and idle timeScale to zero, confirm instances are offline, delete
Slow first requestThe model has to boot before it can runUse popular public models, or pay to keep an instance warm
No reply from supportEmail-only support with no published response timeOne ticket with exact numbers; protect yourself with limits
Requests blocked, credit unusableModeration on the model, and credit is non-refundableTest with small top-ups before committing a budget
Works on one model, fails on anotherInputs and outputs differ per modelRead each model’s schema; verify state after any 500

What Replicate bills, by model type

All quotes are from Replicate’s billing documentation.

What you runSetup timeIdle timeActive timeFailed run
Public modelFreeFreeBilledNot billed
Private model on dedicated hardwareBilledBilledBilledBilled for the instance time
DeploymentBilledBilledBilledBilled for the instance time
Fast-booting fine-tuneFreeFreeBilledNot billed
  • Public models: “you only pay for the time it’s active processing your requests. Setup and idle time for the model is free.”
  • Private models: “you pay for all the time instances of the model are online”.
  • Deployments: “we charge for all the time deployment instances are online.”
  • Failures: “For all models, if a run fails, we don’t charge you”, but for private models and deployments the instance was online during the failed run, and that time is billed. A canceled run on an official model “may still be charged”.

So the same model can be free while waiting as a public model and billed while waiting as a deployment. Many teams create a deployment to avoid cold starts and do not notice that they have also switched billing models.

Four charges people do not expect

1. A deployment that bills while it does nothing

Deployments exist to keep instances warm. If the minimum instance count is above zero, you pay around the clock whether or not traffic arrives. With a minimum of zero you still pay for each boot and for the idle period before an instance scales down.

The harder case is an instance that never finishes booting. In one public report from August 2026, a user describes a private deployment showing 100% setup time and 0% active time, with the balance rising from about $9 to over $91 in 18 hours after they had disabled it and set the maximum instances to zero. This is one user’s account and we cannot verify it, but it shows why you should confirm that instances have actually stopped instead of trusting the setting.

2. Cold starts you pay for

On a public model the boot is free. On a deployment or private model, you pay for the boot as well. Large models take longer to load, so the least used and largest models cost the most per successful request.

3. Credit that expires

Replicate’s prepaid credit page says purchased credit “is valid for 1 year from the purchase date” and “is not refundable”. If you top up a large amount for a project that then pauses, the unused part is gone after a year. The same page says that when the balance reaches zero “we will prevent any new work from starting”, and that in rare cases a prediction can run past the balance, with the difference charged to your payment method at the end of the month.

4. An early charge

The billing documentation says that “if your usage crosses certain limits for the first time, we’ll charge you early”, as a fraud check, and that the amount shows up as a credit on the next invoice. It is not an extra fee, but it is a card charge you did not schedule.

How to stop a charge that is running now

  1. List every deployment and private model. Old experiments are the most common source. Check each one, not only the one you think is the problem.
  2. Set minimum and maximum instances to zero on anything you do not need, then watch the instance count and the metrics panel for a few minutes. The number that matters is instances online, not the value in the settings form.
  3. Delete what you no longer use. Replicate’s docs say “You can only delete deployments that have been offline and unused for at least 15 minutes”, so scale to zero first and delete after that wait. If deletion keeps failing because instances are still online, that is the evidence to send to support.
  4. Revoke API tokens you cannot account for. A leaked token can start work on your account. Create a new token for each service so you can see which one is spending.
  5. Email support with numbers. Include the deployment name, the time you scaled it to zero, a screenshot of the metrics showing setup or idle time after that point, and the balance at two timestamps. A request with exact times is easier to credit than “I was overcharged”.
  6. Turn auto-reload off while you investigate. If the balance is refilled automatically, a runaway instance keeps being funded.

How to keep it from happening again

  • Use public models unless you need a deployment. You get free setup and idle time. Move to a deployment only when cold starts cost you more than idle hardware does, and write that comparison down.
  • Cap maximum instances. Replicate’s deployment docs describe this as the way to limit spend.
  • Buy credit in amounts you will use within the year.
  • Alert on spend, not on errors. A stuck instance produces no errors. Poll your balance or billing usage daily and alert on the change.
  • Review deployments on a schedule. A monthly list of everything online, with an owner next to each item, removes most surprises.

Slow cold starts

A cold start is the wait while a model is loaded onto a machine before your request can run. Replicate’s documentation says a prediction stays in starting while this happens. Popular public models are usually already loaded. Less used models are not, and the first request after a quiet period waits for the whole boot. Even a recent positive review on Trustpilot names cold starts on less popular models as the drawback.

What you can do on Replicate:

  • Do not wait synchronously. Create the prediction, then poll or use a webhook. A blocking call that times out while the status is still starting looks like a failure to your user but is not one.
  • Set a deadline in your own code and tell the user the model is warming up instead of showing a spinner with no end.
  • Keep an instance warm with a deployment if the model is central to your product. This removes the wait and brings back the idle billing described above, so it is a trade, not a fix.
  • Prefer models that are already popular when two models would do the job equally well.

Support that does not reply

Most of the one-star reviews on Replicate’s Trustpilot page are not about the product. They are about not getting an answer: a customer who says requests went unanswered despite substantial monthly spending, another who could not log in or reach anyone for months, and the deployment report above, filed publicly after 18 hours without a reply while the balance kept rising. The page has only 17 reviews, so it is a sample of unhappy users, not a measurement. Still, the pattern is consistent enough to plan around.

  • Send one complete ticket. Account name, the resource, exact timestamps, what you expected, what happened, and the amount in question. Reply in the same thread instead of opening new ones.
  • Do not rely on support to stop a cost. Scale to zero and revoke tokens yourself first; ask for the credit afterwards.
  • Ask about an escalation path before you depend on the platform, not during an incident. If your revenue depends on the API, a named contact is worth asking for.

Blocked requests and credit you cannot use

Some models apply their own content moderation, and a request can be refused for a prompt or an image you consider harmless. One reviewer in October 2026 describes Seedance requests repeatedly rejected with copyright errors while the purchased credit remained on the account and support did not answer.

Two documented facts matter here. Replicate says a failed run is not charged, so a blocked request should not cost you anything. But purchased credit “is not refundable” and expires after a year, so if the model you bought it for keeps refusing your use case, the money stays on the platform.

  • Top up a small amount first and run your real prompts and reference images through the exact model before buying more.
  • Log the full error for each refusal. A moderation refusal is not retryable; an infrastructure error is.
  • Have a second model for the same task, since moderation differs between models.

API behavior that changes from model to model

Replicate hosts models published by many different authors, and each one defines its own inputs and outputs. That flexibility is the point of the platform, and it is also where integrations break.

  • File inputs. In a September 2026 issue, a developer found that several image models rejected valid PNG files with “Invalid image format” until the client was told to send files as base64, which the model pages did not mention. If one model rejects a file that others accept, try the other file encoding before debugging the image.
  • Output shape. Some models return a single file, others a list, others text. Normalize the output in one place in your code.
  • Errors that are not failures. Another report describes a training request that returned a 500 while the training job actually started and completed. After any 5xx on a create call, list recent predictions or trainings before you retry, or you may pay for the same job twice.
  • Versions. Community models are addressed by version id. Pin the version you tested and upgrade on purpose.

If you would rather not manage any of this

Hardware-time billing makes sense when you run your own custom model at high, steady volume. If you mostly call well-known image, video and audio models, a per-request price is simpler to predict.

WaveSpeedAI prices each model per request or per unit of output, such as per image or per second of video, and there is nothing to keep warm:

  • The price is shown before you run. Each page in the model catalog shows the cost of the request you configured.
  • No cold starts. Models are ready when you call them. You do not create deployments, set instance counts or pay for boot and idle time.
  • Failed tasks are refunded in full. If a task fails for any reason, including a blocked request, 100% of its cost goes back to your balance.
  • One API for every model. Every model uses the same two calls, the same authentication and the same result format, and files are passed as URLs. Adding a model means changing the model path and its parameters, not your integration.
  • Support that answers. You can reach the team through the chat on the site or on Discord and get a timely reply from a person.
  • Spending limit per API key. You can cap how much a key may spend, so a test script or a leaked key cannot drain the account.
  • Auto top-up is optional, and a request is simply refused with Insufficient credits. Please top up your account to continue when credits run out.

A request is two calls: submit, then fetch the result.

curl --request POST 'https://api.wavespeed.ai/api/v3/wavespeed-ai/z-image/turbo' \
  --header "Authorization: Bearer $WAVESPEED_API_KEY" \
  --header 'Content-Type: application/json' \
  --data '{"prompt": "A lighthouse on a cliff at sunrise, soft fog, 35mm photo"}'
curl --request GET "https://api.wavespeed.ai/api/v3/predictions/$PREDICTION_ID/result" \
  --header "Authorization: Bearer $WAVESPEED_API_KEY"

The first call returns data.id; the second returns data.status and, when completed, the file URLs in data.outputs.

Where Replicate remains the better fit: pushing your own container with Cog, and the long tail of community models. If your product depends on a custom model you packaged yourself, keep it there and apply the controls above. To try the hosted models without writing code first, use the image generator or the video generator.

For a broader comparison, see Replicate vs WaveSpeedAI.

FAQ

Why is Replicate charging me when I am not running anything?

Most likely a deployment or private model has instances online. Those are billed for setup, idle and active time, not only for predictions. Check every deployment’s instance count, scale unused ones to zero and delete them.

Does Replicate charge for failed predictions?

Replicate says a failed run is not charged. For deployments and private models, however, you pay for the time the instance was online, which includes the time spent on a run that failed.

Does Replicate credit expire?

Yes. The prepaid credit page says purchased credit is valid for one year from the purchase date and is not refundable.

How do I stop a Replicate deployment from billing?

Set minimum and maximum instances to zero, confirm in the metrics that no instance is online, then delete the deployment once it has been offline and unused for at least 15 minutes. If instances stay online after that, contact support with timestamps.

Do I pay for cold starts on Replicate?

Not on public models, where setup time is free. On deployments and private models the boot time is billed like any other instance time.

Why is my Replicate prediction stuck on “starting”?

The model is booting. Less used models are not kept loaded, so the first request waits for a cold start. Poll for the result instead of waiting on one blocking call, or keep an instance warm with a deployment and accept the idle cost.

How do I contact Replicate support, and how long does a reply take?

Support is by email and Replicate does not publish a response time. Send one ticket with exact timestamps and amounts, and stop any running cost yourself before you wait for an answer.

My request was blocked. Do I get the credit back?

A failed run is not charged, so the blocked request itself should cost nothing. Credit you already purchased is not refundable and expires one year after purchase.

Why does the same code work on one Replicate model and fail on another?

Each model defines its own inputs and outputs. Check the model’s schema, and if a valid file is rejected, try sending it with the other file encoding.

Is there a way to pay per image or per second of video instead?

Yes. Platforms that price per request, WaveSpeedAI among them, show a fixed price for each request before you run it, have no idle or setup charges, and refund failed tasks in full.

Share