Deepgram Nova-3 batch $0.26/hour vs streaming $0.29: 500 hours vs Sume

On Deepgram's pay-as-you-go page, Nova-3 monolingual costs $0.0043 per minute pre-recorded and $0.0048 per minute streaming. For 500 hours (30,000 minutes) that is $129 pre-recorded and $144 streaming, a gap of $15, or about 12 percent. Sume STT is $0.01 per audio minute, so 500 hours is $300.

Deepgram's page also lists Whisper Large at $0.0048 per minute pre-recorded. Everything here was read on Deepgram's page and Sume's catalog on 2026-10-09.

The numbers

Deepgram says it bills true per-second with no rounding. The per-hour figures on the page are $0.26 pre-recorded and $0.29 streaming for Nova-3 monolingual. Multiply by 500 for the archive.

Why streaming costs more

A streaming model has to return words while audio is still arriving, which is a harder service to run than reading a finished file. Deepgram prices the two modes separately. If you are transcribing an archive that already exists, there is no reason to send it through the streaming endpoint; the pre-recorded row is the one that matches the job.

A common mistake is to build a pipeline on a streaming example from documentation and keep it for batch work. On 500 hours that mistake is $15 on this price list, which is small, but the same habit on a multilingual archive costs more.

Where Sume fits at 500 hours

Sume STT would cost $300 for 500 hours, more than double the Deepgram pre-recorded row. It also caps a job at 10 minutes of reserved audio, so 500 hours is at least 3,000 jobs. The reason to choose it is not the rate. It is that the audio, the transcript and the next steps, such as burned-in captions or a timeline render, stay inside one account and one set of job routes.

If you only need the transcripts, the cheaper row is the better choice. If transcripts are the first step of a Sume video pipeline, the difference may be worth paying.

• Existing files: use the pre-recorded price.
• Live audio: use streaming.
• Sume pipelines: weigh the $171 gap against the cost of moving files between vendors.

A note on the unit

Deepgram quotes both a per-minute and a per-hour figure. The per-hour numbers on the page are rounded forms of the per-minute ones, so $0.0043 per minute is shown as $0.26 per hour although 60 x $0.0043 is $0.258. This post uses the per-minute rate times the minutes, which is the more exact of the two.

Checking your own mix

Most archives are not 500 hours of one thing. If 400 of your hours are English and 100 are mixed-language, the monolingual pre-recorded rate covers the first group at 24,000 x $0.0043 = $103.20 and the multilingual rate covers the second at 6,000 x $0.0052 = $31.20, for $134.40. Sume's single rate gives $300 regardless.

Compute your own mix before you decide. The point of a price table is to give you the inputs for that arithmetic, not to replace it.