Summary: OpenRouter launched video generation at /api/v1/videos with asynchronous submit, poll, and download; the first wave includes Seedance 2.0/1.5, Veo 3.1, Wan 2.7/2.6, and Sora 2 Pro. Normalized parameters, capability discovery, endpoint routing, and passthrough options are intended to absorb differences in duration, resolution, references, frame control, and billing. The models endpoint is meant to let agents validate parameters before calls, and each model retains its own listed pricing.
OpenRouter has switched on video generation: the leading video models are now reachable through one API, taking their place next to text, images, audio, embeddings, and rerankers — all routed, governed, and billed through the same layer. Text-to-video and image-to-video are both available from day one, with Seedance 2.0/1.5, Veo 3.1, Wan 2.7/2.6, and Sora 2 Pro in the first wave and more models on the way; the new endpoint is /api/v1/videos.
The Case for a Single Layer: How Video APIs Have Fragmented
No two providers describe a video request the same way — request shapes, parameter names, and billing units all diverge. OpenRouter tackles this with four normalization designs: asynchronous generation — a single run can take minutes, so the whole flow is managed as a job: you send a prompt, receive a job ID, and collect the video once it’s ready; normalized parameters — every model accepts the same schema, spanning resolution, duration, aspect ratio, audio generation, frame images, and reference images; capability discovery — query programmatically what each model supports before you call; and passthrough parameters — model-specific features pass straight through when you need them.
A Minefield of Parameters: One Wrong duration Kills the Call
The differences between these models run deeper than what meets the eye. The textbook case is duration: Veo 3.1 accepts 4, 6, or 8-second clips, while Wan 2.6 takes 5 or 10 — supply any other value and the request fails on the spot. Fragmentation also shows up in endpoint families: a single model’s text-to-video, image-to-video, and reference-to-video often live on separate endpoints. OpenRouter reads your parameters and sends the request to the correct endpoint automatically, so that layer of complexity never reaches the caller. How reference images (characters) and frame images (first/last frames) get passed in has been standardized as well.
The Discovery Endpoint: Designed With Coding Agents in Mind
Hit /api/v1/videos/models once and you get back every model’s supported resolutions, aspect ratios, pricing, input images, and durations — OpenRouter itself dubs it “a perfect endpoint to give your coding agent”: holding that data, an agent can adapt itself to each model instead of having to discover valid inputs through error messages. Model-specific features still have an outlet: Veo 3.1’s exclusive personGeneration parameter (which decides whether people show up in the output) is the showcase passthrough case. Model pages also picked up a Playground tab where outputs can be tried directly.
Multimodal Pipelines: An LLM Drafts, an Image Model Casts, a Video Model Films
Nothing excites OpenRouter more, by its own account, than composition: an LLM stretches a rough idea into a full prompt, an image model draws the lead character, and a video model places that character inside a scene — one API the whole way. The practical takeaway is put plainly: these models reward specific descriptions. Camera movement, lighting, texture, pacing, motion style — every one of them shapes the result, and the finer the detail you supply, the more control you hold, which is exactly why video generation pairs so well with LLM-written prompts. OpenRouter has open-sourced a demo app, multimedia-explorer (code on GitHub), that runs this multimodal pipeline from start to finish.
Practical Entry Points
Three places to start: first, hook /api/v1/videos/models into your agent or codegen flow from the outset, leaving parameter validation to capability discovery rather than trial-and-error; second, each model is billed at its own listed pricing, so one read of the manifest covers cross-model cost comparison; third, when the goal is a production-grade multimodal pipeline, lift the architecture from the open-source multimedia-explorer — its LLM→image→video chain already runs end to end. Feedback and requests for the next batch of models belong in the #video-feedback channel on Discord.