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

Replicate prepaid credit: negative balances and auto-reload

Understand Replicate prepaid credit, negative balances and auto-reload. Separate funding from spending controls and reconcile usage before adding more money.

Replicate prepaid credit: negative balances and auto-reload

You bought credit before running the model, so a negative balance seems impossible. Replicate's prepaid documentation nevertheless allows for it: in rare cases, a prediction can cost more than the remaining credit. The outstanding amount is then charged to the default payment method at the end of the month.[1]

That exception sits beside another rule on the same page: at zero, Replicate prevents new work from starting and shuts down infrastructure. The two statements need to be read together. The service has a depletion rule, but the amount prepaid is not an unconditional maximum for every prediction already running.

If you have found a negative balance, stop new submissions from the application you control and preserve the relevant prediction IDs. Then work out which usage crossed the remaining balance. Adding credit first can make the account usable again without explaining what happened.

One prediction can change the meaning of “only spend what's left”

Take a hypothetical experiment with three credit units remaining. A prediction's final cost turns out to be five units. Under the documented overrun exception, the shortfall would be two units. These are invented units for arithmetic, not a Replicate model price or an incident we reproduced.

The question is not whether five is greater than three. It is when you know that the job will cost five. A displayed balance tells you what funding remains at that observation; it is not a quote for the next model run. If the job's charge depends on work performed, the final amount may only be known afterward.

Replicate's billing documentation distinguishes public-model active processing from private-model and deployment costs, which can include setup and idle infrastructure as well as active time.[2] Identify the billing arrangement used by the disputed work before multiplying a request count by an assumed fixed price. Ten requests and ten hours of provisioned infrastructure are different measurements.

Now add a second worker to the hypothetical experiment. Both workers see the same three-unit balance and independently decide that a two-unit estimate fits. Each local comparison passes; their combined estimate is four. This illustrates an application budgeting mistake, not a claim about Replicate's internal admission logic. The shared balance was counted twice by the application.

Checking the provider balance more often does not, by itself, coordinate those decisions. The application also needs to know what its other workers have already committed. An unresolved job should remain visible in that calculation, especially when a local timeout has left its remote outcome unknown.

Find the work behind the shortfall

Start with the account or organization whose balance is negative. Match that account to the credentials and applications involved. Include jobs submitted through the website or a teammate's script if they belong to the same funding pool; a clean log from one worker does not account for the whole account.

Replicate documents both a prediction lookup and a list of predictions associated with the authenticated user or organization.[3] Use those records together with the IDs saved by your application. Keep the model or deployment, submission time and final state beside each disputed job so similar runs do not get merged into one.

For a job still starting or processing, retain its unresolved status. A client timeout describes what the client stopped waiting for; it does not establish the final remote result. Fetch the known prediction's state before treating a retry as necessary. A new submission can create another piece of work to reconcile.

If you cancel a prediction, keep the cancellation result and inspect its eventual state. For a prediction that reaches its deadline, Replicate's lifecycle documentation distinguishes an aborted prediction before processing from a canceled prediction after processing begins; processing already performed can be billable.[4] Cancellation should not be entered as a blanket refund in your own records. Private deployments also need their separate infrastructure charges checked.

Match settled usage to purchases over the period in question. Keep an automatic reload, a manual credit purchase and an outstanding-usage payment as different events. The same card can fund more than one kind of transaction; its statement alone does not explain the corresponding account movement.

If the records still do not add up, send support the account, prediction IDs, purchase references, timestamps and the amount left unexplained. State the discrepancy narrowly. For the hypothetical five-unit job above, the question would be whether that prediction accounts for the two-unit shortfall—not whether prepaid services can ever have outstanding usage. Replicate has already documented that possibility.

Auto-reload buys more funding; it does not approve the workload

Replicate's auto-reload uses a threshold and a target balance. In its documented example, a ten-dollar threshold and fifty-dollar target cause a forty-dollar addition when the balance reaches ten dollars. The target is the funded balance after the reload, not a fifty-dollar purchase every time.[1]

That distinction helps explain a payment, but it does not decide whether the application should keep submitting jobs. A reload can keep an intended batch running. It can also fund an unintended retry loop. Before raising the target to cure repeated interruptions, identify what is consuming the credit.

Saving the settings can itself initiate funding. Replicate says a reload starts immediately if the current balance is at or below the chosen threshold when you save.[1] Read the values before confirming them; the save action is not merely a preview of what might happen later.

Turning auto-reload off changes future automatic funding. Do not record that setting change as proof that earlier usage disappeared. The documentation treats optional reloads and month-end collection of a rare overrun separately. Reconcile the outstanding work regardless of how you choose to fund the next batch.

There is also a reason not to buy an unnecessarily large buffer: Replicate's documented purchased credit expires one year after purchase and is non-refundable.[1] Buy for a workload you expect to run, with those conditions in view. Credit expiry and negative usage are separate questions; an expiry policy does not explain which prediction caused a shortfall.

The spending-limit tutorial may describe an old account

A search for a monthly cap can surface Replicate's older spend-limit announcement. The page now carries a notice that new customers can no longer set spend limits as of July 1, 2025, and directs customers needing the feature to support.[5] An old screenshot is therefore not enough to establish what your account can enforce today.

Check the controls actually available to your account and their scope. If a limit is present, determine whether it covers the relevant work and how already accepted usage is handled. Do not infer a precise stopping guarantee from a label such as “budget” or “limit.”

For a bounded experiment, write the intended decision plainly: how much work is approved, who can authorize more and which submission paths are included. A low-balance notification can prompt that decision, but receiving an alert is not the same as stopping every worker.

This also separates two legitimate goals. A production service may prioritize continuity and allow automatic funding. A one-off evaluation may prioritize a small, reviewed workload and stop submitting when that allowance is committed. Copying the production account's reload settings into the evaluation does not automatically give it the second policy.

Reserve the experiment's budget before submitting its next job

An application can keep its own record of approved spending, settled usage and outstanding commitments. This is a proposed design for your application, not a description of a built-in Replicate feature:

remaining application allowance =
  approved experiment budget
  - settled usage
  - estimated cost of outstanding jobs

Suppose the experiment has ten units approved, four units settled and two pending jobs estimated at two units each. The application has two units left to allocate, even if its current provider balance looks larger. Those numbers are illustrative. The important part is that both pending jobs remain in the calculation.

When one pending job settles at one unit, replace its two-unit estimate with the actual one-unit charge. Do not subtract both the estimate and the final charge; that would count the same job twice. If the other job's outcome is unknown, retain its provisional commitment instead of deleting it because the worker stopped waiting.

Make admission a shared decision across workers. Two workers must not both allocate the same last two units from separate copies of the budget. Include manual submissions in the approved scope or keep them explicitly separate. Otherwise, the application can enforce its own calculation correctly while another route spends outside it.

An estimate remains an estimate. If you cannot establish a reliable upper bound for the chosen workload, run a smaller approved batch and inspect its settled results before expanding it. Calling an uncertain estimate a hard monetary cap would recreate the original misunderstanding in a different dashboard.

Changing providers does not remove that question. reAPI's task reference documents reservation followed by final settlement, including model-specific handling where actual GPT usage exceeds the available balance.[6] A visible balance held at zero is not, by itself, proof that no shortfall exists. Compare the exact contract for the model you intend to use; the Replicate pricing guide covers the broader billing choices.

For the next batch, the decision should be based on the work still outstanding and the allowance still uncommitted. In the two-worker example, that means one shared decision about the remaining funds before either worker submits again. Another balance refresh cannot make the same three units available twice.

References

  1. Replicate. Prepaid credit: exhaustion, negative balances, auto-reload and purchase conditions. Retrieved November 15, 2026.
  2. Replicate. Billing: public models, private models and deployments. Retrieved November 15, 2026.
  3. Replicate. HTTP API: get and list predictions. Retrieved November 15, 2026.
  4. Replicate. Prediction lifecycle: cancellation and deadlines. Retrieved November 15, 2026.
  5. Replicate. Set a monthly spend limit. Historical changelog with deprecation notice, checked November 15, 2026.
  6. reAPI. Tasks: reservation and final settlement. Read November 15, 2026.
All resources