Live vs file transcription cost: gpt-live-transcribe and batch rates
Streaming transcription costs 2 to 4 times file transcription on vendor pages. When live is worth it, and when Sume's $0.01 per minute file job is enough.
If the audio already exists as a file, pay for file transcription. The vendor pages price streaming at about two to four times the batch rate: OpenAI's gpt-live-transcribe is $0.017 per minute against $0.0045 for gpt-transcribe, and xAI's streaming STT is $0.20 per hour against $0.10 for REST. Pay the live premium only when words must appear while someone is still talking.
The rates
| Vendor | Live | File | Ratio |
|---|---|---|---|
| OpenAI | gpt-live-transcribe $0.017 per min | gpt-transcribe $0.0045 per min | about 3.8x |
| xAI | $0.20 per hour | $0.10 per hour | 2x |
| Sume | not offered | $0.01 per audio minute | no live |
When live is worth it
- Live captions on a call or stream.
- A voice agent that must react mid-sentence.
- Accessibility features where a delay of several seconds breaks the experience.
When a file job is enough
- Podcasts, meetings and webinars after they end.
- Subtitle files and searchable archives.
- Cutting video by transcript. Sume returns word-level start and end times with every result, so you can cut on words. See a clickable transcript from word timestamps.
Worked example
For a 45-minute recording, OpenAI file transcription is 45 x $0.0045 = $0.2025, OpenAI live is 45 x $0.017 = $0.765, and Sume is 45 x $0.01 = $0.45, split into five requests because the cap is 10 minutes each. The Sume figure assumes you send an accurate duration_seconds for each clip.
Limits
Sume does not offer streaming transcription. Rates here are list prices and exclude any plan discounts, and no accuracy comparison is made. See the full file-rate table in cheapest speech-to-text per hour.
Sources
Related posts
More in Comparisons
- LiveKit lists 16 avatar providers: real-time vs rendered Sume clips
LiveKit Agents lists 16 avatar providers that join a call as a participant. Sume Avatar 1.0 is not one of them; it renders finished clips. When to use which.
- Longest single AI video clip, October 2026: vendor limits compared
Kling 4.0 lists 3 to 30 seconds, Luma Ray3.2 up to 20, Gemini Omni Flash 3 to 10. A read-dated table of the stated limits and the Sume cap beside them.
- Loudly music API tempo, genre and mood vs a Sume Music prompt
Loudly's music API exposes mood, tempo and genre as inputs. Sume's Music Router has one prompt field and no tempo parameter, so BPM goes in the text.
- Luma Dream Machine add_audio endpoint vs Sume generate_audio
Luma lets you POST /generations/{id}/audio to add sound to a finished clip. Sume sets audio at submit with generate_audio; how to get the same result.
Written by Sume