OpenRouter adds ByteDance's Seedance 2.5 for 30-second video generation

Why it matters

Seedance 2.5 moves AI video toward reference-heavy editing and production workflows. For OpenRouter, it tests whether a unified API remains valuable when access comes from one provider and routing is unavailable.

Alex Atallah, the OpenSea co-founder who now runs OpenRouter, added ByteDance's Seedance 2.5 to the gateway's video-generation API on August 7th, giving developers a single endpoint for clips lasting up to 30 seconds and prompts containing dozens of reference assets.

https://x.com/OpenRouter/status/2086260287683092792

OpenRouter said in a thread on X that one request can include as many as 30 images, 10 video clips and 10 audio clips. Those inputs can guide the model's subjects, composition, movement and sound, or be used to request edits such as replacing a person and adding, removing or modifying objects in an existing video.

The release extends Atallah's original routing thesis into one of the most technically uneven parts of generative AI. Atallah previously co-founded and served as chief technology officer of OpenSea after working at Palantir and several earlier startups. He started OpenRouter in 2023 as a neutral interface across competing AI models, aiming to spare developers from writing and maintaining separate integrations for each provider.

A larger reference window

ByteDance's Seed research unit describes Seedance 2.5 as an audio-video model designed for 30-second storytelling, reference-controlled generation and editing. ByteDance says a clip can be extended twice, while the model also supports camera controls, performance blocking and green-screen editing intended for production workflows.

The reference limit is a substantial increase from Seedance 2.0, which ByteDance released on February 12th. That version accepted up to nine images, three video clips and three audio clips, then produced up to 15 seconds of synchronized video and audio. Seedance 2.5 doubles the native output duration while lifting the combined reference allowance from 15 assets to 50, according to the specifications published by OpenRouter and ByteDance.

That increase matters for applications that need to preserve a character, product or visual identity across a sequence. A developer could supply separate references for a subject, location, camera treatment, soundtrack and source footage in one job instead of compressing those instructions into a text prompt or stitching together several independent generations.

OpenRouter also lists first-frame and first-and-last-frame controls, optional generated audio, multilingual audiovisual generation, video extension and reference-based editing. OpenRouter describes Seedance 2.5 as suited to long-form storytelling, although its native 30-second ceiling still requires extensions or post-production assembly for longer work.

Developers access the model through OpenRouter's asynchronous video API. An application submits a job to the video endpoint, receives an identifier and polling URL, waits for generation to finish, and then downloads the completed file. OpenRouter also supports webhooks for applications that do not want to poll each job repeatedly.

The model page lists pricing from $0.1028 per generated second. Its provider table, however, shows $0.2311 per second for both video-only and video-with-audio output through Seed. Developers should inspect the active endpoint price before estimating production costs because the headline rate and displayed provider rate do not match.

Seed is currently the only provider listed for Seedance 2.5 on OpenRouter. OpenRouter therefore forwards each request directly to that endpoint and cannot choose among competing hosts for price, latency or availability. The immediate benefit is API standardization rather than the provider-level failover that OpenRouter offers when several endpoints serve the same model.

OpenRouter pushes beyond text models

Video broadens the role OpenRouter wants to occupy between application developers and model creators. OpenRouter began as a marketplace for language models, but its current interface also covers image, audio, transcription, embedding and video workloads. OpenRouter says it now handles more than 200 trillion tokens a month across more than 400 models from over 70 providers and serves more than 10 million users. Those figures are self-reported.

The addition comes less than three months after OpenRouter raised a $113 million Series B led by CapitalG. NVentures, ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures, AMP PBC and Pace Capital joined the May 28th financing alongside existing investors Andreessen Horowitz and Menlo Ventures. OpenRouter said the capital would fund infrastructure, enterprise controls and more sophisticated routing.

Seedance 2.5 also exposes the boundary of that routing pitch. OpenRouter can give developers one billing relationship, request format and job-management system for video models, reducing the engineering work required to test or replace them. With only one Seedance 2.5 provider available, OpenRouter cannot create competition beneath the request or recover by sending it to another host. That could change if additional providers begin serving compatible versions of the model.

For Atallah, the release is another attempt to make the gateway valuable even when OpenRouter does not control the underlying model or compute. Model makers are releasing new modalities, input formats and pricing schemes on different schedules. Each additional integration increases the maintenance burden for application developers and gives OpenRouter another reason to sit between those developers and the labs.

Seedance 2.5 gives that strategy a demanding test case. Video jobs are slower, more expensive and operationally different from text inference. They require asynchronous queues, file storage, progress tracking and large multimodal uploads. If OpenRouter can make those differences look like another model selection inside one API, Atallah's router becomes harder to remove from the application stack even when it has only one upstream endpoint to offer.