视频生成功能现已上线 OpenRouter。通过单一 API，您即可访问顶尖视频模型。视频现已与文本、图像、音频、嵌入向量及重排序模型并列，统一纳入相同的路由、治理与计费体系。
上线首日，我们支持 Seedance 2.0 / 1.5、Veo 3.1、Wan 2.7 / 2.6 及 Sora 2 Pro 的文本转视频与图像转视频功能，后续还将接入更多模型。浏览所有受支持的模型，了解更多功能详情，或直接跳转至新的 /api/v1/videos API 文档。
视频 API 目前较为分散。各提供商使用不同的请求格式、参数名称和计费单位。我们正是围绕这些差异构建了该 API：
- 异步生成：此类生成任务耗时数分钟，因此我们将其作为任务进行追踪。提交提示词后，获取任务 ID，待视频就绪后即可检索。
- 标准化参数：我们提供一套适用于所有模型的统一模式，涵盖分辨率、时长、宽高比、音频生成、帧图像及参考图像。
- 能力发现：在调用模型之前，可通过编程方式确定每个模型支持的功能。
- 透传参数：在需要时，可直接使用模型特有的功能参数。
统一 API
视频模型通常以端点族的形式发布。每个端点提供特定能力，例如文本转视频、图像转视频或参考图像转视频。我们通过根据您的参数自动路由至正确端点来简化这一过程。
视频模型之间的差异并非总是显而易见。即使是时长设置也容易导致请求失败。Veo 3.1 支持 4、6 或 8 秒的片段，而 Wan 2.6 支持 5 或 10 秒的片段。我们还标准化了参考图像（例如角色）和帧图像（例如首帧和末帧）的传入方式。
每个模型还可暴露其自身特性。例如，Veo 3.1 包含一个独特的 personGeneration 参数，用于控制输出中是否出现人物。/api/v1/videos/models 端点会准确告知您哪些模型特有参数可用。
无需自行追踪所有这些差异，只需调用视频模型端点，即可在一个位置查看支持的解析度、宽高比、定价、输入图像和时长：`/api/v1/videos/models`。这是一个非常适合提供给编码智能体的端点，它能提供所有必要细节，让智能体适应每个模型，而无需因可接受的输入参数报错而反复调试。
为了提供视觉预览，我们还在模型页面上新增了一个“Playground”选项卡，方便您试用每个模型并查看其生成效果。
多模态工作流
我们最兴奋的是看到将视频与其他类型的生成内容相结合所取得的成果。通过同一个 API，大语言模型可以将一个粗略的想法转化为详细的提示词，图像模型可以生成主角，而视频模型则可以将该角色转化为一个场景。
我们很快认识到，这些模型对细节的精确度要求很高。镜头运动、光照、纹理、节奏、动作风格——所有这些都很重要。您提供的细节越多，获得的控制力就越强。这使得视频生成成为大语言模型生成提示词的自然应用场景。
我们构建了一个开源演示应用，展示了这一多模态工作流的实际效果：multimedia-explorer.openrouter.ai。所有代码均托管在 GitHub 上。
请在 Discord 的 #video-feedback 频道告诉我们您的想法以及您希望接下来支持哪些模型。感谢所有参与 Alpha 测试并帮助塑造 API 的朋友。如果您想抢先体验后续功能，请加入我们的 Discord 社区。
Video generation is now live on OpenRouter. One API gives you access to the top video models. Video now sits alongside text, images, audio, embeddings, and rerankers, under the same routing, governance, and billing layer.
On day one, we’re supporting text-to-video and image-to-video on Seedance 2.0 / 1.5, Veo 3.1, Wan 2.7 / 2.6, and Sora 2 Pro, with many more to come. Browse all supported models, learn more about the feature, or jump straight to the API docs for the new /api/v1/videos.
Video APIs are fragmented. Providers use different request shapes, parameter names, and billing units. We built the API around those differences:
- Asynchronous generation: These generations take minutes, so we track them as jobs. Submit a prompt, get a job ID, and retrieve the video when ready.
- Normalized parameters: We provide one schema that works across every model, including resolution, duration, aspect ratio, audio gen, frame images, and reference images.
- Capability discovery: Programmatically determine what each model supports before you call it.
- Passthrough parameters: Use model-specific features directly when you need them.
One Unified API
It’s typical for video models to be released as a family of endpoints. Each endpoint provides a specific capability, such as text-to-video, image-to-video, or reference-to-video. We simplify this by automatically routing you to the correct endpoint based on your parameters.
Video models vary in ways that aren’t always obvious. Even duration can easily break a request. Veo 3.1 supports 4, 6, or 8-second clips, while Wan 2.6 supports 5 or 10-second clips. We also standardize how you pass in reference images (e.g., characters) and frame images (e.g., the first and last frames).
Each model can also expose its own features. Veo 3.1, for example, includes a unique personGeneration parameter that controls whether people appear in the output. The /api/v1/videos/models endpoint tells you exactly which model-specific parameters are available.
Instead of tracking all these differences yourself, you can call the video models endpoint to inspect supported resolutions, aspect ratios, pricing, input images, and durations in one place: /api/v1/videos/models. This is a perfect endpoint to give your coding agent, providing all the details it needs to adapt to each model without battling errors over acceptable input params.
For visual previews, we also added a new Playground tab on model pages, so you can try each model and see what it can make.
Multimodal Workflows
We’ve been most excited to see the results when combining video alongside other types of generations. An LLM can turn a rough idea into a detailed prompt, an image model can generate a main character, and a video model can turn that character into a scene, all through one API.
We’ve learned quickly that these models reward specificity. Camera movement, lighting, texture, pacing, motion style, all of it matters. The more detail you provide, the more control you get. That makes video generation a natural fit for LLM-generated prompts.
We built an open-source demo app that shows this multi-modal workflow in action: multimedia-explorer.openrouter.ai. All the code is on GitHub.
Tell us what you think and which models you want next in #video-feedback on Discord. Thanks to everyone who tested the alpha and helped shape the API. If you want early access to what’s next, join us on Discord.