GPT Image 2.5 is live — OpenAI's newest image model, targeted edits that leave the rest of the frame alone
All resources

· reAPI Team

AI API charges with no requests: idle deployment costs

Trace AI API charges when no requests are running. Check dedicated endpoints, idle instances, retained storage and the evidence that a deployment has stopped.

AI API charges with no requests: idle deployment costs

You have sent no prompts and generated no videos, yet another charge appears on the account.

Before you investigate a mystery API call, look at the unit on the bill. If it measures online hardware time or model hosting, an empty request history may be entirely consistent with the charge. Together AI bills active dedicated endpoints for allocated hardware and deployed fine-tunes for hosting. Replicate bills deployment instances while they are setting up, processing and idle.[1][2]

That difference changes what you need to turn off. Closing the app stops you using it. Stopping a deployed endpoint changes the resource that generates the bill.

One hour of work, ten hours online

Suppose you deploy a model for an experiment, use it for an hour, then leave the instance online for the rest of a ten-hour window. For this example, assume the price is two cost units per online instance-hour. These are invented units for arithmetic, not a quote from any provider.

The bill is twenty units: one instance multiplied by ten online hours, multiplied by two units. Counting only the hour of useful processing gives two units. The remaining eighteen are not unexplained requests. They are the cost of keeping the instance online during the other nine hours, under the billing assumption in this example.

Now change the experiment. You send half as many prompts but keep the same instance online for the same ten hours. The hosting total is still twenty units. Shorter prompts and fewer output tokens do not change this particular meter. You would need to change the billable online time, number of instances or applicable hardware rate.

For a real deployment, use its documented billing unit and actual instance intervals, including any setup time and rounding the service applies. Replicate's billing documentation explicitly includes setup, idle and active time for deployments. Public models use a different billing model, and fast-booting fine-tunes have an exception; do not transfer a rule from one execution type to another just because the model names look similar.[2]

This also explains why an API can appear expensive when you divide the bill by a small number of completed requests. You are spreading the cost of keeping capacity ready over the work that used it. Whether that is worthwhile depends on why you kept it ready.

What the idle instance buys you

Replicate's deployment controls let you set a minimum instance count to keep models warm, or let unused instances scale down to zero. They also let you cap the maximum number of instances.[3] The minimum and maximum answer different questions. A low maximum restricts how far the deployment can grow; a nonzero minimum can keep capacity online while no new work arrives.

For an interactive application, keeping an instance ready may be a reasonable choice. The next request does not have to wait for that instance to be prepared from an offline state. For an occasional internal experiment, paying through long quiet periods may be harder to justify. The documentation establishes the controls, but it cannot tell you how much startup delay your particular users will accept.

Make that decision against the workload. A batch that can wait until the next scheduled run has a different requirement from a user watching a live interface. If you have measured startup behavior in your own approved tests, use those measurements. Without them, you cannot yet tell whether the additional waiting time fits your application's needs. Keeping an instance warm does not establish how quickly every request will complete, either.

The same distinction applies at Together. Its support guidance separates serverless model usage from deployed fine-tune hosting and dedicated endpoints. Serverless usage is charged for actual API usage; the hosting and dedicated resources can continue accruing charges without it.[1] “We use Together” is not enough detail to choose a billing rule. Find the endpoint type on the account.

A useful note beside an experiment's deployment is therefore its reason for staying online: the workload it serves, who owns it, and when that need ends. It makes a later shutdown decision easier than trying to infer the purpose of an old endpoint from its name.

Find the resource before removing the payment card

Start with one bill line and its service period. Identify the resource ID and account or organization, then open the corresponding endpoint or deployment. Compare its state and allocated instances with the period being charged. The question is whether that resource was billable during that interval.

If the bill gives you only a total deduction, look for the usage detail behind it. Do not compare a hardware-hour entry with the model's token price. A credit purchase is a funding event, while hosting consumes that funding; Together's support guidance explicitly separates the two.[1]

Next, check that you have the full set of resources in the relevant account. Include the development endpoint, the deployment from last month's evaluation and resources created by other team members. You do not need to assume any of them was forgotten. You need an owner who can say whether each should still be running.

If the charge is for inference rather than hosting, compare provider request history with the application's logs. Check scheduled jobs and workers as possible sources of activity. A closed browser window is weak evidence about a service running elsewhere. If you find requests you cannot explain, investigate those separately; the leaked-key response guide covers credential exposure.

Also check when accepted jobs finished. reAPI documents that pending tasks can complete and consume credits after their API key is revoked. Its task response distinguishes a reservation from final settlement.[4] In that situation, a later balance change can belong to earlier work. Compare submission and completion times before concluding that a new request ran after shutdown.

Removing a payment method does not answer any of these questions. Together requires at least one valid card and a replacement default before the current default can be removed.[5] Its documented way to stop unused hosting is to stop the endpoint. Deal with the billable resource first; otherwise you can spend time changing payment settings while leaving the relevant endpoint active.

Stop it, then inspect the next billing interval

For Together, the published procedure is to review deployed fine-tunes and active dedicated endpoints, then explicitly stop the ones you no longer use.[1] Record which endpoint you stopped and when. If the team still needs the model, keep its configuration and deployment requirements available for the next person who has to bring it back.

For Replicate, inspect both the scaling configuration and actual instance state. A setting describes what the deployment is allowed to do; monitoring tells you whether instances are starting, idle or processing.[3] After a change, check the resulting online count. Do not use a saved configuration screen as your only evidence that the resource is offline.

Deletion is a further step. Replicate's documentation requires a deployment to have been offline and unused for at least fifteen minutes before it can be deleted.[6] If your immediate goal is to stop paying for an abandoned experiment, follow the lifecycle in that order. Decide separately whether the team wants to retain its configuration or remove the deployment.

Then look at the next usage record. Compare the interval it covers with your shutdown time, not just the date the charge appeared. A line posted after shutdown can still describe earlier billable time. A genuinely new billable interval after the resource became inactive is a different finding, and one worth taking to support with the resource ID and timestamps.

Return to the ten-hour example. If the record says the instance was online for all ten hours, the twenty-unit total follows the stated assumption. If you have evidence that it went offline earlier, the disagreement is now precise: which interval did the billing system count? That is a much more useful support question than “I did not send any prompts.”

A stopped model can leave other costs behind

Turning off compute does not necessarily remove the resources around it. AWS provides a concrete example: EBS volumes persist when an EC2 instance stops and continue to incur storage charges.[7] This is an AWS storage rule, not a claim that Together or Replicate uses the same storage fees.

If you operate your own infrastructure, read the remaining bill items individually. Identify the retained volume or other separately billed resource and decide what it contains before removing it. AWS documents that deleting an EBS volume removes its data, with snapshots providing a separate retention option.[8] Saving money is not a reason to delete the only copy of a model artifact the team still needs.

The same care applies to outputs you have copied into your own storage. The generation service's lifecycle does not decide the lifecycle of your archive. The output retention guide follows those copies through saving and deletion. For a shutdown checklist, give storage its own owner and decision rather than hiding it under “endpoint stopped.”

Scaling one deployment to zero therefore cannot establish that the whole account now costs nothing. There may be other endpoints, earlier work settling, or separately billed resources. Close out each item against its own unit and service period.

Decide whether you need to own a deployment at all

Once the bill makes sense, ask what the application needs to run next time. Custom weights, infrastructure control and dedicated capacity are reasons to evaluate a deployment. Calling a supported, ready-made model may instead fit a hosted model API. Those are different product choices, not interchangeable ways of buying the same thing.

For supported model inference, compare the reAPI catalog with the model and input/output contract your app requires. The task documentation shows how results and settled usage are reported; the balance reference covers the account side.[4] This does not make reAPI a replacement for custom training or dedicated GPU hosting. If those are requirements, include them in the comparison rather than pricing only the easiest part to replace.

There is no single useful cost to “create an AI app” without that choice. Estimate the work to build it separately from inference, online compute, storage and operations. For a lightly used app, a small request total can coexist with a substantial hosting total, as the example above shows. For a busy service, keeping capacity ready may be intentional.

Free allowances do not resolve the architecture question either. Replicate documents selected models that can be tried before billing setup, while OpenAI documents prepaid API credits that decrease with paid usage.[2][9] Neither establishes unlimited free inference or free dedicated hosting. When the next experiment is approved, record how it is billed and who will stop it when the experiment ends.

References

  1. Together AI. Why am I being charged when I'm not making API calls? Retrieved November 15, 2026.
  2. Replicate. Billing. Retrieved November 15, 2026.
  3. Replicate. Deployments. Retrieved November 15, 2026.
  4. reAPI. Tasks, Balance and Authentication references. Retrieved November 15, 2026. Task lifecycle, balance, and authentication.
  5. Together AI. Can I add multiple payment cards? Retrieved November 15, 2026.
  6. Replicate. Delete a deployment. Retrieved November 15, 2026.
  7. Amazon Web Services. How EC2 instance stop and start works. Retrieved November 15, 2026. Instance lifecycle and related costs.
  8. Amazon Web Services. Delete an Amazon EBS volume. Retrieved November 15, 2026. Volume deletion and retention.
  9. OpenAI. Setting up and managing prepaid API billing. Retrieved November 15, 2026. Prepaid API billing.
All resources