Reserving AI minutes: the accounting behind AI Clips

AI Clips uses processing minutes, so clicking Generate is an accounting event as well as a product action. Creator includes 60 minutes per month, Scale includes 600, and Enterprise is effectively unlimited. Free and Hobby do not include AI Clips.

The service reserves minutes before sending work to the remote processor. If it waited until the end, two clicks at the same time could spend the same balance twice. If it charged every failed request, a creator could lose minutes without receiving clips. The rules are therefore: reserve on dispatch, refund known failures, and allow one job per recording.

Reserve before starting the job

When you generate clips, ApexStream rounds the recording duration up to whole minutes. A 61-second recording uses two minutes. The request is clamped to zero or higher so a bad duration cannot create a credit.

The usage row is updated conditionally: the reservation succeeds only when the current usage plus the request is within the monthly limit. That update is atomic, so two simultaneous requests cannot both pass the same stale balance. Only after the reservation succeeds does the service create a short-lived source URL and the remote project.

If project creation fails because of a network error, rejection, or missing source path, the reserved minutes are refunded immediately. There is no useful reason to hold a reservation for work that never became a job.

The reservation is based on the recording duration, not on how many highlights eventually appear. A long recording can therefore consume more of the monthly pool even if the result contains only a few clips. The clip-count allowance and the minute ledger answer different questions.

One job per recording

A recording already marked PROCESSING cannot receive another simultaneous AI Clips request. That avoids three competing progress states and three reservations for the same source.

After a hard failure, you can try again. A retry is new work, so it creates a new reservation. The dashboard does not provide a cancellation path that automatically returns minutes while a remote job is still running; the poller owns the lifecycle.

Known failure means a refund

A background poller watches the remote project. When it reports a known failure—such as corrupt video, an explicit failure code, or a project that disappeared—the recording is marked failed and the original duration reservation is refunded.

When the job succeeds, ApexStream saves the clip titles, scores, time ranges, and thumbnails and marks the recording complete. The minutes remain consumed because the processing ran. The number of clips shown per stream is a separate plan rule: Creator shows 3 clips per stream, while Scale and Enterprise are unlimited.

A successful run can still require editorial judgment. Scores and time ranges help you find candidate moments, but they do not decide whether a clip has the context, rights, or tone you want to publish. The product gives you downloadable MP4 results; it does not auto-post them to social platforms.

The timeout case

A remote job can also stop making progress without returning a normal failure. If it hangs for around an hour, the system may mark it failed without passing through the same known-failure refund path. In that case, a reservation can remain consumed even though clips were not produced.

That is a documented edge between the poller, queue retries, and a remote worker that has gone quiet. If a failed, non-producing job appears to have taken minutes, contact support@apexstream.co for human review. The honest explanation is better than promising automatic recovery that the lifecycle does not always provide.

What this means when you use AI Clips

Think of the monthly minutes as a shared budget for the account, not a per-recording allowance. A short clip request and a long recording request do not have the same cost, even though both appear as one Generate action.

• Make sure the recording plays before spending minutes on it.
• Do not press Generate repeatedly; a recording allows one in-flight job.
• Remember that a retry consumes a new reservation after a hard failure.
• Watch the monthly pool: Creator 60 minutes, Scale 600, Enterprise uncapped.

The interface should feel simple. The ledger underneath needs to be strict enough that concurrent clicks and failed jobs have a clear owner.

Reserve first, run the work, refund failures we can identify, and describe the timeout seam when we cannot. That is the accounting model behind the one-click experience.

Frequently asked questions

When are AI Clip minutes used?

They are reserved when ApexStream processes a selected recording for AI Clips. They are not the same as the number of clips you eventually download.

What happens when the allowance is gone?

The feature should show the remaining allowance or the plan gate. Keep the source recording and use the transcript editor or manual clip workflow when AI processing is not available.