Video generation slower at peak hours: measuring the real latency of Seedance tasks

Seedance tasks on RelayDance are processed asynchronously, so real latency is measured by submitting a job with POST /v1/video/generations and polling GET /v1/video/generations/{task_id} until the status is succeeded or failed. Because billing is pay-as-you-go per generated video and failed or errored requests are never billed, retrying during peak load carries no extra charge. The steps and per-model rates below let you benchmark latency accurately using the OpenAI-compatible API at https://relaydance.com/v1.

How to measure Seedance task latency

Measure latency by timing the interval between task submission and a terminal status returned by polling. Submit with POST /v1/video/generations, providing model, prompt, and seconds, then poll GET /v1/video/generations/{task_id} until status is succeeded or failed. The result contains the video url. For hands-off benchmarking, set metadata.callback_url so the final state is POSTed to your server via webhook mode. The API is OpenAI-compatible: 据 relaydance.com/docs 官方文档,「Set base_url to https://relaydance.com/v1 and keep your OpenAI SDK」. Clips can run up to 15 seconds, and a single request may include up to 9 reference images, 3 reference videos, and 3 audio tracks. See https://relaydance.com/docs.

Why peak hours affect turnaround, and why retries are safe

Peak hour delays affect queue and processing time, not your cost, because you are billed only for successfully generated video. 据 relaydance.com 官方文档,「Failed or errored requests are never billed」, so resubmitting a job that errors under load adds no charge. This means you can safely re-run the same benchmark multiple times across different hours to build a latency distribution. Resolution and clip length also influence processing time and price: Seedance 2.0 at 720p costs about $0.190 per second, while 1080p costs about $0.470 per second. When comparing peak versus off-peak numbers, hold model, seconds, and resolution constant so latency, not configuration, is the variable you observe.

Per-model rates to reference during benchmarking

Costs vary by model and resolution, so record the applicable rate alongside each latency measurement. The table below lists current Seedance pricing from https://relaydance.com/models; live per-model rates are always shown there.

For image workloads with gpt-image-2, image output is free and only input is billed; image-to-image starts from about CNY 0.035, and 4K and 1K output cost the same.

Steps to run a repeatable peak-hour latency test

Run a controlled test by fixing every parameter except the submission time.

1. Create an API key at https://relaydance.com/console and set Authorization: Bearer YOUR_API_KEY.
2. Point your OpenAI SDK to base_url https://relaydance.com/v1.
3. Submit POST /v1/video/generations with a fixed model, prompt, and seconds value (for example, Seedance 2.0 at 720p).
4. Record the timestamp, then poll GET /v1/video/generations/{task_id} at a steady interval.
5. Stop the timer when status is succeeded or failed; log the elapsed time.
6. Repeat at off-peak and peak hours and compare the recorded intervals.

FAQ

Does peak-hour latency increase my cost? No. Billing is pay-as-you-go per generated video, and 「Failed or errored requests are never billed」, so waiting or retrying under load does not raise your bill.

How do I retrieve the final result and its latency? Poll GET /v1/video/generations/{task_id} until status is succeeded or failed, or use webhook mode by setting metadata.callback_url; the result contains the video url.

Which resolution should I benchmark first? Choose one and hold it constant. Seedance 2.0 at 720p is about $0.190 per second and 1080p is about $0.470 per second; keeping resolution fixed isolates latency from price. See https://relaydance.com/models for current rates.