API Cost and Quotas
Billing units shape the real cost. Estimate how many runs a task takes before comparing headline rates.
Related
Cheap per call, costly per keeper
A thirty-second cue that needs eight generations before one is usable costs eight times the list rate, plus the time spent repairing stems. Compare billing units first, then count how many runs a task really takes. Comparing headline rates alone usually leads to the wrong choice.
Three billing bases
| Basis | Charged on | What raises the bill | Good for | Easy to miss |
|---|---|---|---|---|
| Per second | Input or output duration | Length, sample rate, channels | Scores, effects, long beds | Silence and fade tails still count |
| Per character | Prompt, lyrics, text | Word count, retries, context | Lyrics, descriptions, bulk copy | System prompts and chat history |
| Per task | One generation or edit | Candidates, resolution, concurrency | Full songs, stems, mastering | Failed jobs, queuing, timeouts |
Some services blend the three: a monthly allowance, then per-second or per-character overage. Allowances that expire monthly are a quiet price increase. Per-task pricing looks simple until candidate count and retry policy multiply the total.
Estimating one job
Cost ≈ unit rate × billable amount per run × number of attempts + repair hours + storage and transfer.
An example: a two-minute instrumental billed on output duration, six generations per keeper on average, two of them discarded for structural errors. Count eight runs, not one. If stems must be exported and cleaned, labour often exceeds the API charge. Keep the formula in billing units rather than currency so it survives a change of vendor.
A checklist of hidden costs
- Whether failed, timed-out or cancelled jobs are billed
- Listening and selection time across candidates
- Extra charges for stems, MIDI or project-file export
- Storage, download traffic and long-term archiving
- Queueing caused by concurrency caps
- Exchange rates, tax and regional pricing
- Retesting after terms change
When metered billing loses
High-frequency small jobs such as stingers, loop batches and teaching demos are usually steadier on a subscription or a local setup. A handful of long cues a year is cheaper metered. Low unit price is not low total price. Put quota, concurrency and rework into the same comparison, and keep a redo column in the budget, or the first failed job will distort the whole sheet.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- 生成服务公开的 API 文档，关于计费单位、配额与错误码的说明。
- 云计算计费白皮书中关于存储、流量与请求计费的通用做法。
- 后期制作工时估算的行业惯例，关于分轨修整与试听筛选。