Video generation is now live on OpenRouter: one API for the top video models, sitting alongside text, images, audio, embeddings, and rerankers under the same routing, governance, and billing layer. Day one covers text-to-video and image-to-video across Seedance 2.0/1.5, Veo 3.1, Wan 2.7/2.6, and Sora 2 Pro, with more to come; the new endpoint is /api/v1/videos.
Why Unification: The Fragmentation of Video APIs
Every provider ships a different request shape, parameter names, and billing units. OpenRouter’s answer rests on four normalization designs: asynchronous generation (these generations take minutes, so everything is tracked as a job — submit a prompt, get a job ID, retrieve the video when ready); normalized parameters (one schema across every model, covering resolution, duration, aspect ratio, audio generation, frame images, and reference images); capability discovery (programmatically determine what each model supports before you call); and passthrough parameters (use model-specific features directly when you need them).
The Parameter Minefield: Even Duration Breaks Requests
Video models differ in ways that are not obvious. The canonical example is duration: Veo 3.1 supports 4, 6, or 8-second clips while Wan 2.6 supports 5 or 10 — the wrong value fails the request outright. Another layer of fragmentation is endpoint families: the same model often splits text-to-video, image-to-video, and reference-to-video across separate endpoints. OpenRouter routes you to the correct endpoint automatically based on your parameters, hiding that complexity from the caller. How you pass reference images (characters) and frame images (first/last frames) is standardized too.
The Capability Endpoint: A Perfect Interface for Coding Agents
A single call to /api/v1/videos/models returns each model’s supported resolutions, aspect ratios, pricing, input images, and durations — and OpenRouter names it explicitly as “a perfect endpoint to give your coding agent”: the agent adapts to each model from that manifest instead of battling errors over acceptable inputs. Model-specific features have an escape hatch: Veo 3.1’s unique personGeneration parameter (controlling whether people appear in the output) is the passthrough example. Model pages also gained a Playground tab for trying outputs directly.
Multimodal Workflows: LLM Prompts, Image Characters, Video Scenes
The application OpenRouter says it is most excited about is composition: an LLM expands a rough idea into a detailed prompt, an image model generates the main character, and a video model turns that character into a scene — all through one API. The hands-on lesson is blunt: these models reward specificity. Camera movement, lighting, texture, pacing, motion style — all of it matters, and the more detail you provide, the more control you get, which makes video generation a natural fit for LLM-generated prompts. OpenRouter open-sourced a demo app, multimedia-explorer (code on GitHub), that walks this multimodal pipeline end to end.
Adoption Notes
Three practical starting points: wire /api/v1/videos/models into your agent or codegen flow first, and let capability discovery handle parameter validation instead of trial-and-error; billing follows each model’s pricing, and cross-model comparison fits in a single manifest read; and for a production multimodal pipeline, borrow the architecture from the open-source multimedia-explorer — the LLM→image→video chain already works there. Feedback and requests for the next models go to the #video-feedback channel on Discord.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.