Every job records what it actually cost, with a breakdown by stage, at the rates set in the settings. The settings page shows the month's spend against the budget.
Estimates and actual costs
- Before queueing, the file's page and the batch screen's confirmation show an estimate: the recording's duration (from ffprobe, or from the bit rate Omeka read when the file was added) at the transcription rate, plus an estimate of the tokens of the language-model passes.
- After each paid stage, the actual cost is recorded from the provider's answer: the audio duration it reports, and the tokens it used.
| Rate | Default | Applies to |
|---|---|---|
| Transcription rate (USD a minute) | 0.006 | Minutes of audio sent for transcription. OpenAI whisper-1. |
| Model input rate (USD per million tokens) | 0.15 | Prompt tokens of the cleanup, segmentation and summary passes. |
| Model output rate (USD per million tokens) | 0.60 | Completion tokens of the same passes. |
Change the rates when your provider's prices change; they apply to estimates and to later actual costs, not to past records. Costs are kept in whole cents: a short clip can cost $0.00.
Caps
| Cap | Behavior |
|---|---|
| Monthly budget (USD) | Queueing is refused when the month's spend (UTC) plus the estimate would go over it, and a running job stops before its next paid stage once the spend is over it, failing with budget_exceeded. Retry it when there is room again (next month, or with a higher budget). 0 means no budget. |
| Daily automatic queue cap | At most this many files are queued automatically a day (UTC). Files passed over are listed on the review queue page; staff queueing is not limited. |
Keeping costs down
- A retry resumes at the stage that failed, so a passing provider error never bills completed stages again.
- Re-run segmentation and Generate summary each run one pass.
- Leave Summary element empty if you do not want summaries: the pass is skipped.
- The mock provider costs nothing, and is good for training staff on the review workflow.