gpt-realtime-whisper $0.017 per minute vs Sume STT file jobs

OpenAI's model page lists gpt-realtime-whisper at $0.017 per minute, served through the realtime transcription sessions endpoint. Sume STT 1.0 costs $0.01 per audio minute but only transcribes a finished file at a public HTTPS URL. If you need words while someone is still speaking, Sume is not the tool; if you transcribe recordings, the lower per-minute rate and the job model apply.

What the OpenAI page lists

The page states the price, the endpoint and per-tier rate limits from 100 to 1,300 audio minutes per minute, which matters for live products with many simultaneous callers.

The Sume side

A Sume STT request returns text and words[] with timings once the job completes. You poll GET /v1/jobs/:id/status or use a webhook, then read /result.

import os, requests

r = requests.post(
    "https://api.sume.com/v1/stt-1.0/transcribe",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}",
             "Idempotency-Key": "stt-demo-001"},
    json={"audio_url": "https://example.com/call.mp3",
          "language_code": "en",
          "duration_seconds": 300},
)
print(r.status_code, r.json())

Decision rule

Use streaming for voice agents, live captions and call coaching. Use a file job for recordings, podcast archives, and the transcript step before captions or dubbing. Mixing both is normal: stream during the call, and send the saved recording for a cleaner final transcript.

Worked example

A support line with 100 hours of calls per month is a common size to test.

• 100 hours at $0.017 per minute: streaming transcription $102.00.
• The same 100 hours as recordings on Sume STT: $60.00.
• Streaming and file jobs are different products: only streaming returns words while the caller is still talking.

Checklist before you commit

Many teams use both: a live stream for agent assist and a recorded-file pass for the archive, which is where a file job fits.

• Write down your latency target before choosing.
• Check OpenAI's rate-limit tier for your account against peak concurrent calls.
• Keep call recordings only where your policies permit.