Seedance 2.5 Flash joins Vercel's AI Gateway catalog for cheaper AI video
ByteDance launched a lighter version of Seedance 2.5, and Vercel has already made the model available in the AI Gateway, integrated with AI SDK 7. For those building generative video features, this means less latency and lower cost per clip, without switching providers.

- Reading
- 4 min
- Code
- 2
ByteDance launched a lighter version of Seedance 2.5, and Vercel has already made the model available in the AI Gateway, integrated with AI SDK 7. For those building generative video features, this means less latency and lower cost per clip, without switching providers.
Vercel added bytedance/seedance-2.5-flash to the AI Gateway catalog, according to an announcement published in the platform's official changelog. The model is a faster, cheaper version of Seedance 2.5, by ByteDance, and arrives integrated with AI SDK 7 and the AI CLI. For those who already use the Gateway to orchestrate text and image models, video now becomes part of the same workflow, with the same unified API.
What changes between Seedance 2.5 and Flash
According to Vercel, Flash delivers output "close to the standard model in human realism, reference-based editing and prompt adherence," but with lower cost and lower latency. In practice, it's the model designed to run in production, where every second of generation and every cent per clip factor into the product's bottom line. It's the same logic as other "flash/pro" pairs that already exist for text and image models, now applied to video.
The model generates clips from 4 to 30 seconds long, at up to 720p and 24 frames per second, in landscape, square and portrait formats. It also supports editing and extending existing videos, in addition to allowing first and last frame images to be defined to control the start and end of the scene. Audio comes out synchronized with the video, generated from the same prompt.
How to call the model with AI SDK 7
The integration uses the experimental experimental_generateVideo function, renamed in the code as generateVideo. The example published by Vercel shows the simplest call, with text only:
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,
});The call uses the same AI Gateway already configured in the project, within AI SDK 7. Those who already have the Gateway configured get video without needing to add a new integration.
Image, video and audio references in the prompt
The most interesting point for those building a product is the support for references. The inputReferences parameter accepts images and videos as context, and the prompt references each one by number: [Image 1], [Video 1], and so on. Vercel's example illustrates the use case well:
const { videos } = await generateVideo({
model: 'bytedance/seedance-2.5-flash',
prompt: 'The lion from [Image 1] walks through a sunlit meadow, filmed like [Video 1].',
aspectRatio: '16:9',
duration: 5,
inputReferences: [
{ data: '...', mediaType: 'image/jpeg' },
{ data: '...', mediaType: 'video/mp4' },
],
});This opens the door to features like "animate this character" from a photo uploaded by the user, or "recreate this camera style" from a reference clip. It's a different pattern from pure text-to-image: the model is combining multiple input media into a single coherent generation, which demands more care in prompt design and error handling when the reference isn't well interpreted.
What this replaces (and what it doesn't)
For those who already integrated ByteDance's API directly, going through the AI Gateway trades a proprietary call for a standardized AI SDK interface. In short: the gain isn't just price, it's code portability between generative video providers, using the same AI SDK interface to switch models without rewriting the integration.
This doesn't replace, however, a traditional video editing pipeline. Seedance 2.5 Flash generates short clips controlled by prompt or reference; it doesn't cut, doesn't assemble and doesn't do color grading on a larger project. Those who need full post-production will still need to integrate the output of these models into an editing tool afterward, whether via code or manually.
When it makes sense to swap the standard model for Flash
Vercel did not publish pricing or latency numbers in the changelog, so the decision between standard Seedance 2.5 and Flash depends on testing both models in the Gateway's own playground before deciding. Some practical criteria to guide that test:
- Real-time product feature (instant preview, generation within an onboarding flow): Flash tends to make more sense due to lower latency.
- Marketing material or hero video, where quality matters more than speed: it's worth comparing side by side with the standard model before setting Flash as the default.
- High volume of generations (thousands of clips per day): Flash's per-unit cost changes the product's financial viability, even with a small loss of fidelity.
What remains open
Vercel's changelog does not detail rate limits, price per second of generated video, or availability SLA for Flash compared to the standard model, information that normally appears in the Gateway's pricing documentation, not in the announcement. There is also no objective quality comparison between the two versions in the source, only the claim that the result is "close" to the standard model. Anyone deciding between the two models in production needs to run their own A/B test before switching the project's default, using the video generation quickstart and the model playground that Vercel itself provides for this comparison.
Translated from the Brazilian Portuguese original · Read the original
Vercel's AI Gateway unifies private models and catalog models in a single API
The AI Gateway's Private Inference feature lets you plug fine-tuned, self-hosted, or prerelease endpoints into the same authentication, SDKs, and observability used for Vercel's catalog models.





