What $50 buys on Sume: speech, transcripts, images, music, video

Fifty dollars of Sume balance buys 1,052,631 characters of speech, 83 hours of transcription, 2,222 background removals, 400 Music 1.0 tracks, or 400 seconds of Wan 3.0 video at 720p, if you spend it all on one thing. Each figure divides $50 by the rate on the API pricing page as read on 2026-10-05, and rounds down because you cannot buy a fraction of a job.

The table

Every Sume price is the provider list times 1.25, and one workspace balance pays for all of it, as described in the billing docs. The table shows what $50 buys of a single thing at a time. Real spending is a mix, so treat it as a sense of scale.

What the numbers mean

The spread is the point. Speech and transcription are cheap per unit and cost cents per thousand of them, music is cheap per track, and video is where $50 disappears: 400 seconds is under seven minutes of Wan footage at 720p and 200 seconds at 1080p. Seedance 2.5 costs more per second again, so the same $50 buys far less of it.

For a project that is mostly video, the other lines are rounding error. That is why budgets for short-video work should be built from seconds of footage first, with narration, music and captions added afterwards as small line items rather than the other way round. A 30-second 720p Wan clip is $3.75, and the narration for it, say 900 characters, is $0.043.

A split that looks like real work

Here is one way to spend $50 on a batch of ten short videos. Each has 30 seconds of Wan 720p ($3.75), a 900-character voiceover ($0.04275), one music track ($0.125), a one-minute timeline render ($0.10) and a caption job ($0.20). That is $4.21775 per video and $42.18 for ten, leaving about $7.82 for retries.

• Video: $37.50 for ten videos, 89 percent of the total.
• Captions: $2.00 for ten.
• Music, voice and render together: $2.68 for ten.

Limits on the figures

Some jobs reserve more than they capture, and Sume's docs say a request reserves its price up front, so a balance has to cover the reservation when you submit. Your actual spend, job by job, is visible in the usage docs surface. A single request also has limits: TTS takes 20,000 characters, STT 10 minutes, and Wan 2 to 30 seconds. Seedance 2.5, Kling and the image models with higher tiers cost more per unit than the rows above, so a plan built on them will land well short of these quantities. Both limits and prices are listed per model in the video docs and image docs. The rows are meant for orders of magnitude, not for a promise about a particular model.

How to run your own

Divide your budget by the rate. For a different budget, multiply the quantities above: $200 buys four times as much, and $500 ten times. A useful habit is to price a batch of ten before a batch of a thousand, since the ten-unit cost is a quick sanity check on the arithmetic and on your assumptions about duration or length. Check the live rate for each before you commit, because the numbers here are only as current as the pricing page on the day it was read.