· reAPI Team
fal.ai failed requests billing: what is and isn't charged
Understand fal.ai failed requests billing, including server errors, invalid inputs and timeouts. Reconcile task IDs before retrying or requesting a refund.

A charge next to a failed image feels like paying for nothing. fal's published policy makes a narrower distinction: server errors at HTTP 500 and above are not charged, but an invalid-input error at HTTP 422 may be charged if a runner used GPU time before detecting the problem.[1]
That is the starting point for checking a bill. “Failed” in your app is not yet a billing category. You need the response from the particular request, and you need to know whether the failure happened during generation or while your app was trying to retrieve the result.
The difference between a 500 and a 422
For fal Model APIs, the official FAQ says infrastructure failures do not cost you anything. It gives HTTP 500 and above as the server-error category. If your request record shows that outcome and a corresponding settled charge, you have a specific policy discrepancy to ask support about.[1]
Check where the response came from before making that comparison. A 500 from your own application while it displays an image is not the same record as a 500 returned by the model request. Keep the failing URL or operation beside the error. Otherwise the same three digits can lead you to apply the generation policy to a later step.
A 422 needs a closer look. fal says invalid inputs may be detected after the runner has already spent GPU time; that work can be billable.[1] The policy does not say every 422 is charged. Nor does the phrase “invalid input” establish that the service rejected the request before doing any work.
Read the message and the selected model's input requirements together. If a required field or file property caused the failure, correct that condition before submitting a replacement. Avoid changing several unrelated settings at once: you will make it harder to tell which change addressed the documented error.
Content-policy failures also need their actual classification. fal's FAQ identifies content_policy_violation as a non-retryable 422 error.[1] Repeatedly sending the same rejected input is not a useful infrastructure recovery strategy. For billing, retain the error and the request's settled usage; the error name alone does not tell you how much was ultimately charged.
Free credits can fund a trial without changing the underlying distinction. fal documents grants and coupons with their own validity conditions.[1] Check what your account actually has; do not assume that every user gets free usage or that a replacement attempt will be free. Plan for paid model usage once the applicable grant no longer covers it.
The fal.ai pricing guide covers the broader cost model. Here, a model's advertised starting price is less useful than the record of the attempt you are disputing.
An error on the screen can belong to a different step
Consider a hypothetical image job. Your application submits it, waits, then times out. You press Generate again. Later you find a lower balance and one image in the interface.
There are now two submissions to investigate. The screen does not establish whether the first failed remotely, whether both completed, or which image you are seeing. Before asking for a refund, find the original request IDs and inspect their outcomes separately. If one completed, try to retrieve that result rather than treating the missing preview as proof that the generation never happened.
This distinction is explicit in WaveSpeed's Sync Mode refund guidance. A timeout in the client, proxy, browser, gateway or network can leave the prediction running on the server. If it completes successfully, that client timeout is not automatically refundable; the guidance tells users to retrieve the prediction by ID.[2] This is WaveSpeed's documented policy, not an extension of fal's 500 rule to another service.
For fal's queue flow, COMPLETED with no displayed result needs its own result-retrieval check. Keep the submission response, status response and result response distinct in your logs. If a later download fails, preserve the URL you received and investigate that file access separately.
When the submission response itself was lost, you have less information. Do not label the original job refunded simply because your code never saved an ID. Check the account history and available request logs, or ask support to locate the event from the model and timestamp. Until then, describe the outcome as unresolved. A new POST can create another attempt without resolving the first one.
Cancelled does not necessarily mean free
Cancellation introduces a timing question. Replicate documents that a prediction reaching its deadline before processing starts is aborted and not charged. If the deadline arrives after processing has begun, the prediction is cancelled and the processing time is billable.[3]
That difference is easy to lose in an app with one Stop button. The user can stop watching, the application can request cancellation, and the provider can confirm a terminal state at different points. Record the confirmed outcome. Closing a tab is not evidence that the provider received or completed a cancellation request.
Replicate's billing rules also distinguish time-priced predictions from official models and deployed infrastructure. A cancelled time-priced run is billed for processing time already used; cancellation of an official model may still be charged. Most private models and deployments have their own instance-time rules.[4] A cancellation policy for one of these does not automatically describe the others.
If you are comparing providers, keep those service types visible. A public model request and a dedicated deployment can incur costs under different meters even when both produce images. The idle deployment cost guide explains why an empty request history does not settle an infrastructure bill.
For the disputed request, the useful questions stay concrete: did processing begin, which terminal state was recorded, and what does that model's billing arrangement charge for it? The amount of media shown to the user cannot answer those on its own.
Put the request and the money on the same line
You do not need a large reporting system to investigate one charge. Start with the provider request ID, model, submission time and time zone. Add the component that reported the failure, the exact error and the latest remote outcome. Then attach the amount you are questioning and any later reversal.
If you retried, give each replacement its own line. Link it to the original job, but do not merge the attempts into one “failed request” entry. In the hypothetical timeout example, that separation is what lets you discover whether the balance change belongs to the first submission, the replacement, or both.
Also distinguish a reservation from settled usage when the platform exposes both. reAPI reserves credits at submission and reports settlement through the task lifecycle. A processing task's zero settled usage does not mean the reservation has been returned. A failed task with zero settled credits is a refunded outcome; some model-specific post-generation safety failures can retain a positive charge.[5]
Those distinctions belong in the investigation even if the dashboard uses a single balance figure. An unavailable amount is not zero. An earlier deduction is not necessarily the final amount. Preserve what the provider actually exposes rather than filling a missing field with the outcome you expect.
A support message can then be short: identify the request, state its recorded final error and settled amount, and name the policy that appears inconsistent with them. Ask support to confirm which billing category applies to that request. Include the related replacement IDs so they can separate repeated work from a duplicate charge.
Remove API keys, card details and unrelated private inputs before sharing the record. Support needs enough context to locate and understand the event, not a public copy of every file in the job. Keep any later correction attached to the same request so the issue can be closed against evidence.
A correct bill can still be an expensive workflow
Reconciling the charge answers whether the policy was applied. It does not answer whether the workflow is good value. A successful output that cannot be used still belongs in the cost of delivering the project, even if the generation service fulfilled its contract.
For that second question, use the total settled generation cost for a workload divided by the number of outputs accepted for use. This is a suggested operating metric, not a provider billing formula. Define what accepted means before evaluating the results: the required dimensions, consistency of the subject and suitability for the intended use, for example.
Keep unreviewed outputs separate from rejected ones. If no output has been accepted, do not report a finite cost per accepted output. You have spent money without reaching that outcome yet. The request-level records still matter because they show which part of the spend came from billable failures, deliberate retakes or unnecessary replacement submissions.
That breakdown tells you where a change might help. An input problem calls for checking the documented requirements. A completed job lost by your app calls for better retrieval. A cancellation that arrives after processing begins calls for reviewing when the application asks to stop. Switching models may be worth evaluating for output suitability, but it is not a substitute for finding out which of those problems you have.
Take a small hypothetical batch: four attempts at one cost unit each, with all four amounts confirmed as settled. Two outputs are accepted, one completed output misses the brief, and one failed attempt is billable under its applicable policy. The spend is four units for two accepted outputs, or two units per accepted output. Neither the unit rate nor those outcomes are measurements from a provider.
Now suppose an investigation establishes that the failed attempt should not have been billed, and one unit is actually returned. The corrected cost is three units divided by two accepted outputs, or one and a half units each. The unusable but successfully generated image still belongs in the project cost. A refund fixed one charge; it did not turn that other image into a deliverable.
References
- fal.ai. Model APIs FAQ: failed requests, credit grants and billing. Retrieved November 15, 2026.
- WaveSpeedAI. Refund Policy: Sync Mode timeouts. Retrieved November 15, 2026.
- Replicate. Prediction lifecycle: cancellation and deadlines. Retrieved November 15, 2026.
- Replicate. Billing: public models, private models, deployments and canceled predictions. Retrieved November 15, 2026.
- reAPI. Tasks: status, reservations and final settlement. Read November 15, 2026.