A usable five-second product video can take many attempts. If you publish Reels, ads, or demos, Vercel's November 16, 2026 announcement gives you another route: ByteDance Seedance 2.5 Flash is now callable through Vercel AI Gateway with a Vercel key. You do not need a separate provider account. The useful decision this week is whether a short clip with synchronized audio and references beats your current production workflow on cost and iteration count.
What actually launched
Vercel added bytedance/seedance-2.5-flash to its video catalog. Flash is positioned as faster and cheaper than standard Seedance 2.5; that is the vendor's description, not a SoyJB benchmark. It takes text plus image, video, or audio references. It can also edit or extend video and use opening and closing frame images. The model page lists clips from 4 to 30 seconds at 480p or 720p, 24 fps, in landscape, square, or portrait.
For a creator, this could mean using an image you own to guide a product shot instead of repeatedly describing the same look. For a founder, the first experiment should be a rough benefit demo, not a finished paid ad.
Access and price
The model is listed as available through AI Gateway. Its current pricing page shows $7 per million generated output tokens. That is not $7 per video: the charge for one clip depends on token use and settings. The playground bills your team at API rates. Vercel says unpaid free accounts get $5 of credit every 30 days; check whether your own account qualifies before relying on it.
Actual provider availability can vary by team or region, and the model page checks it dynamically. Confirm the model appears in your account and set a small spending limit before producing many variants. We have not run our own quality, latency, or per-clip cost test.
A one-hour creator test
- Write one hypothesis: “a moving close-up explains this feature better than a still image.”
- In the playground, choose a five-second portrait clip at 480p. Use an image you own or have licensed. Avoid identifiable people unless you have permission.
- Make one text-only version and one reference-guided version with the same message and length.
- Record the actual charge, wait time, product continuity, readability, and whether you would publish either clip. Only rerender the better candidate at 720p.

Starter prompt: Close-up of an open notebook on a maker's desk. A hand places a yellow card reading “One idea, one test.” Gentle camera movement, natural light, five seconds, vertical framing. Quiet room sound; no spoken voice or invented logos. If the model mangles text, remove it from the generation and add it later in your editor.
Minimal API example for builders
import { experimental_generateVideo as generateVideo } from 'ai';
const { videos } = await generateVideo({
model: 'bytedance/seedance-2.5-flash',
prompt: 'A paper boat drifts down a rain-soaked city street at dusk.',
aspectRatio: '16:9',
duration: 5,
});
This follows Vercel's published example. Configure AI_GATEWAY_API_KEY and a compatible AI SDK first. We did not execute this call or generate a video for this article. For portrait output, select the documented value in the playground or in your installed SDK version rather than assuming a parameter will be accepted.
Where the model may fall short
Flash tops out at 720p, so do not confuse it with the standard model's higher-resolution claims. Review faces, motion, audio, and product details in each output. A reference file does not grant usage rights: only upload media you can use and check the ByteDance terms linked on the model page. Generated text and objects can drift across frames, which matters especially for paid ads and product claims.
Decision: run one five-second reference-guided clip and record cost, delay, and errors. If it is not better than your current workflow, keep that result and skip the integration. Sources: Vercel announcement, November 16, 2026; model details, checked November 17.
