Reference

Limits and retention

Source length caps, rate limits, input sizes, and what happens to retained audio and artifacts.

Source limits

Limit Value
Maximum source length for transcription 30 minutes
Translation input 100,000 characters, 2,000 segments
Summary input 100,000 characters, 2,000 segments
Supported sources Public YouTube videos only

A source longer than 30 minutes fails at quote time rather than after a partial charge. Split longer material into segments before submitting.

Rate limits

Limits are per credential, over a one-minute window. Responses include RateLimit-* headers.

Credential Requests per minute
Account API key 10
Application key 100
OAuth access token Same limit as the underlying account

Account registration is separately limited to five attempts per hour.

Exceeding a limit returns 429 with a retry hint. Back off rather than retrying immediately.

Retention

Artifact Where it lives Notes
Transcript, translation, summary, subtitles Object storage Shared cache
Public source audio Object storage Temporary working copy
Account and credit ledger Managed database Backed up daily
API keys Hashed at rest Returned once, never retrievable

Two consequences worth planning around:

  • Artifacts are shared. A video transcribed by one account is a cached read for anyone else. Do not submit private or sensitive material, because retention is not per-account.
  • Reads stay cheap. Because artifacts persist, a repeat request is a $0.002 read rather than a fresh $0.01-per-minute job.

Idempotency

Spending requests accept an Idempotency-Key. Keys are limited to 128 characters. The same key with the same body returns the original result; the same key with a different body returns 409.

Always send a key on a spending request you might retry.

Deletion

Artifacts are cache entries, not personal storage, so deletion is not exposed as an API call. Treat anything you submit as something that may be retained and reused.