Media pricing is quoted per image, per second, per generation and per character — four units that do not convert into each other on a pricing page. So here is the same exercise as our text-call analysis, run on media: every paid generation-of-one-kind we served in 30 days, aggregated by modality.
| Modality | Calls | Median | p90 | Max |
|---|---|---|---|---|
| Image — Seedream family | 79 | $0.034 | $0.046 | $0.092 |
| Image — WAND Vega | 44 | $0.069 | $0.146 | $0.146 |
| Video | 121 | $0.308 | $0.462 | $1.231 |
| 3D | 17 | $0.277 | $0.369 | $0.369 |
| Speech (TTS) | 75 | $0.0008 | $0.0019 | $0.004 |
Median video clip $0.308 versus $0.034 for a Seedream image. The multiplier is not arbitrary: a 5-second clip is 120+ frames of generation, and video models are billed per second of output (or per token, which scales the same way).
73 paid speech calls cost $0.28 in total. Text-to-speech is billed per character rather than per GPU-second, which is why it sits roughly 40x below an image and 380x below a video clip. If you are adding voice to a product, the cost is not the constraint — latency and voice quality are.
Median $0.277 per 3D generation, close to a video clip. Large-mesh or texture-heavy generations push toward the $0.37 p90.
Seedream list prices start at $0.027 per image, but the measured median is $0.034 — because real requests use higher resolutions and pro tiers. Plan with measured medians, not the cheapest row of the pricing table.
| Workload | At measured median |
|---|---|
| 1,000 images | $34 |
| 1,000 video clips (5s) | $308 |
| 1,000 3D models | $277 |
| 1,000 speech clips | $0.80 |
For a marketing video workflow — say 200 clips a month — that is about $62 at the median. The same budget buys 1,800 images.
Median $0.308 per generated clip across 121 paid video calls in 30 days, with a 90th percentile of $0.462 and a maximum of $1.231. Price depends on model, resolution and clip length.
Median $0.034 for Seedream-family images and $0.069 for WAND Vega images. List prices start lower (about $0.0215) for the cheapest model at base resolution; the measured median is higher because real requests use higher resolutions and pro tiers.
Speech is billed per character, not per second of GPU compute. 73 paid speech calls cost $0.28 in total — a median of $0.0008 per call.
In this sample they are close: median $0.277 per 3D generation versus $0.308 per video clip. But the 3D sample is small (17 calls), so treat it as directional.
Every model per image, per second and per token, from one endpoint with 0% markup. Free tier to test, $2 credit on signup.
Get an API keyDistributions measured from NovAI usage logs over the 30-day window ending 2026-10-22 (336 paid media generations; free-tier calls excluded). Per-model list prices at /china-ai-leaderboard.json.