Cost to transcribe 1,000 hours of audio with an API

Transcribing 1,000 hours of audio is 60,000 minutes. At Sume's $0.01 per audio minute that is $600 before the 5.5% agent fee, and because each request takes at most 10 minutes, it is at least 6,000 jobs. OpenAI's pricing page lists gpt-transcribe at $0.0045 per minute ($270 for the same audio), Whisper at $0.006 ($360) and gpt-live-transcribe at $0.017 ($1,020).

OpenAI figures are from its pricing page, read 2026-10-01. Sume figures are from the rate card and the STT request schema in the API reference. OpenAI's deprecations page lists Whisper-era models for shutdown on February 26, 2027, so budget with the replacement rates.

How many API calls is 1,000 hours?

Sume's duration_seconds hint is capped at 600, and the request schema states a 10-minute maximum. 1,000 hours of audio in 10-minute pieces is 6,000 jobs. Split on silence or at fixed 600-second boundaries, and pass each piece's real length so the reserve matches the work.

What does the whole run cost?

The cost scales linearly with minutes. Sume reserves 1 minute if you omit the hint, so a missing hint on a 10-minute piece under-reserves; send it.

How do I send thousands of jobs safely?

Give every piece its own Idempotency-Key so a retry never double-bills, and send results to a webhook_url instead of polling. Admission limits how many generations run at once by plan (Free 1, Pro 4, Startup 8, Scale 20), and extra requests wait in a queue or fail with a 429 when it is full, so feed jobs steadily rather than in one burst.

curl -X POST https://api.sume.com/v1/stt-1.0/transcribe \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: archive-hour-0001-part-01" \
  -d '{
    "audio_url": "https://media.sume.com/artifacts/artf_demo/part01.wav",
    "duration_seconds": 600,
    "mode": "webhook",
    "webhook_url": "https://example.com/hooks/sume"
  }'