Every pipeline job records what it actually cost, with a per-stage breakdown, using the rates configured in settings.
Estimates and actuals
- Before queueing, the batch screen shows an estimate based on the recording's duration (ffprobe) and the transcription rate, plus a token estimate for the LLM passes.
- After each stage, the actual cost is recorded from the provider's response. The settings page shows the current month's spend.
| Rate | Default | Applies to |
|---|---|---|
| Transcription rate (USD per minute) | 0.006 | Audio minutes sent to the transcription provider. Matches OpenAI whisper-1. |
| LLM input rate (USD per 1M tokens) | 0.15 | Prompt tokens for cleanup, segmentation and summary. |
| LLM output rate (USD per 1M tokens) | 0.60 | Completion tokens for the same passes. |
Update the rates when your provider's pricing changes; they affect estimates and future actuals, not past records.
Caps
| Cap | Scope | Behaviour |
|---|---|---|
| Monthly budget (USD) | Global | Queueing is refused and running jobs stop once the month's spend would exceed it. 0 means unlimited. This cap is authoritative. |
| Daily auto-queue cap | Global | Limits how many media auto-queue per day. Manual queueing is unaffected. |
| Monthly transcription budget (USD) | Per site | Approximate. Cost is attributed to a site only when queueing is site-scoped, which admin-side queueing is not. |
Keeping costs down
- Retries resume at the failed stage, so a transient provider error never re-bills completed stages.
- Re-run segment and Generate/Regenerate summary run one pass each.
- Leave Summary metadata property empty if you do not need summaries; the pass is skipped entirely.
- The mock provider costs nothing and is ideal for training staff on the review workflow.