What changed
Vercel has added asynchronous video generation to AI Gateway. Previously, generateVideo could keep one HTTP request open until rendering finished, which is a poor fit for serverless functions when a job takes seconds or minutes.
The new flow gives builders several execution choices:
startVideoreturns an operation immediately so a queue, serverless function, or parallel job can check it later withgetVideoStatus.generateVideocan wait for a completion webhook instead of holding the original provider request open.generateVideocan poll on a short interval when the calling process can stay alive but cannot receive webhooks.- The Workflow SDK can suspend a durable run while the video renders, then resume it when AI Gateway delivers the completion event.
Existing synchronous calls continue to work, and the asynchronous options cover text-to-video, image-to-video, reference-to-video, and related video inputs. See Vercel's announcement and implementation examples and the video generation documentation.
Why it matters
Video generation is not just a model call; it is a long-running job with timeouts, retries, status, and an eventual artifact. Treating it as a background operation lets an application acknowledge work quickly, persist the operation, and choose a completion path that matches its runtime. That is the difference between a demo that waits on a request and a product that can reliably queue many generations.
The workflow option is especially useful for product teams that already model work as durable steps. Instead of building a custom job table, webhook receiver, and retry loop around every video provider, the workflow can pause without consuming compute and resume when the result is ready. Vercel's Workflow SDK documentation shows the broader runtime pattern.
Builder takeaway
Design the user experience around submit -> status -> artifact, not request -> bytes. Return a job identifier as soon as the generation is accepted, show a clear pending state, and make completion idempotent. If you use startVideo, store the JSON-serializable operation and fetch the result later; if you use webhooks, verify delivery and keep a durable mapping from the job to the user-facing request.
The broader product signal is that AI media APIs are becoming job systems. As generation gets longer and more expensive, the winning abstraction is not only access to a better model, but a reliable way to run, observe, retry, and deliver the output inside a real application.