Runway Media Router 是首个面向生成式媒体的偏好优化路由器,内置于 Runway Dev 平台。
今天,我们推出 Runway Media Router,这是全球首个面向生成式媒体模型的模型路由器。Model Router 是 Runway Dev 中的一项新能力,可为每个请求自动选择最佳的视频、图像或音频模型。在此之前,媒体领域一直没有与 LLM 智能路由相当的能力,主要是因为视频、图像和音频模型在实际能生成的内容上差异极大。我们的 Media Router 通过将请求与模型的能力和模态进行匹配,然后在符合条件的模型之间,根据你设定的成本、质量和延迟偏好进行优化,从而解决了这一问题。
过去,每一次生成请求都意味着要手动挑选模型,并且每当有新模型发布、定价变动或你切换到不同媒介时,都要手动更新这一选择。Model Router 免去了这些步骤,让你只需为你的用例配置一次什么才算“最佳”,并让你在新模型发布时轻松评估和启用它们。
手动挑选媒体模型的问题
Runway Dev 的核心价值,是让用户在一个企业级平台上获取全球最优秀的媒体模型。这意味着既能第一方直接调用 Runway 的前沿模型,包括 Gen-4.5、Aleph 2.0 和 Act-Two,也能使用最优秀的第三方模型,包括 Seedance、GPT Image 2 和 ElevenLabs。
但每一个新模型都在第零天就上线,也成了一种摩擦的来源。开发者告诉我们,他们不想变成兼职的模型研究员,去追踪哪个模型最快、哪个最适合某种风格,并在模型目录每次变动时重新纠结这个选择。我们经常看到团队一旦选定某个模型,就会一直沿用下去,哪怕早已过了它还是正确选择的时点。
在企业规模下,这一差距会不断放大。一个稍贵或稍慢的模型,在单次请求上或许无关紧要,但乘以数千次生成之后,就会变成实实在在的成本与延迟拖累,阻碍规模化。智能路由在企业级 LLM 使用中已相当成熟,但我们认为它对生成式媒体更为重要,因为在这里,质量是一个远比语言复杂得多的维度。
Model Router 如何工作
Model Router 通过三个步骤取代手动选择模型:
设置你的偏好。在 Runway Dev 门户中,在成本、延迟和质量之间选择优化偏好。为你的配置命名,设置价格上限,允许或拒绝不同的模型或提供商,并在你的偏好之间进行优化。将其保存为一次可复用的配置,绑定到特定工作流,比如快速的应内预览与付费的最终导出。
调用一个端点。 将你的生成请求发送到 Model Router 端点,并附上你的配置 ID。无需指定模型。
获得最佳模型。 路由器会过滤掉任何不满足你硬性约束(能力匹配、允许与拒绝列表、价格上限)的模型,然后根据你的成本、质量和延迟偏好对其余模型进行评分。它使用得分最高的模型进行生成,并返回结果以及准确标明使用了哪个模型及其原因的元数据。
今天将推出的初始配置维度包括:
- 价格上限: 每次生成的硬性上限,视频以美元/秒计,图像模型以美元/张计
- 允许与拒绝列表: 提供商和模型级别的规则,将路由限制在已批准的选项内
- 设置偏好: 定义你在成本、质量和延迟优化方面的偏好
如果某个请求在你的约束下无法由任何模型满足,Model Router 会返回一个明确的错误,指明是哪项约束清空了候选池,而不是悄悄降低你的设置。在将配置投入生产之前,一个 dry-run 标志让你可以验证它究竟会选择哪个模型及其原因,且不产生任何费用,也不生成任何内容。
Model Router 随你的团队一同扩展
规模会改变模型选择实际付出的代价:体现在美元、延迟和监管上。
对于请求量很高的团队来说,微小的差异会迅速累积。一分钱的零头或几毫秒的延迟,乘以每年一百万次请求,就会变成一笔可观的成本。Model Router 还解决了一个真实的集成问题:无需不断更新代码,就能跟上快速迭代的模型目录。
立即试用 Model Router
Model Router 现已在 Runway Dev 中面向视频、图像和音频模型上线。前往 dev.runwayml.com 开始构建。

Runway Media Router is the first preference-optimized router for generative media, built into the Runway Dev platform.
Today, we're launching Runway Media Router, the world’s first model router for generative media models. Model Router is a new capability in Runway Dev that automatically selects the best video, image or audio model for every request. Until now, media has never had an equivalent to intelligent routing for LLMs, largely because video, image and audio models vary so much in what they can actually produce. Our Media Router solves this by matching a request to a model’s capabilities and modalities, then optimizing across the qualifying models for the cost, quality and latency preferences you set.
Historically, every generation request meant hand-picking a model and manually updating that choice every time a new model launched, pricing shifted or you switched to a different medium. Model Router removes those steps, letting you configure what "best" means for your use case once, and making it easy for you to evaluate and enable new models as they are released.
The Problem With Picking Media Models Manually
Runway Dev's core value is giving users access to the world's best media models on one enterprise-grade platform. That means first-party access to Runway's frontier models, including Gen-4.5, Aleph 2.0 and Act-Two, alongside the best third-party models, including Seedance, GPT Image 2 and ElevenLabs. But having every new model on day zero has also become a source of friction. Builders tell us they don't want to become part-time model researchers, tracking which model is fastest or best suited to a given style, and re-litigating that choice every time the catalog changes. We often see teams pick a model once and stick with it long past the point it's the right call.
At enterprise scale, this gap compounds. A model that's a bit pricier or a bit slower may not matter on a single request, but multiplied across thousands of generations, it becomes a real cost and latency drag that blocks scaling. Smart routing is well established for enterprise LLM usage, but we think it matters even more for generative media, where quality is a far more complex dimension than it is for language.
How Model Router Works
Model Router replaces manual model selection with three steps:
Set your preferences. In the Runway Dev portal, select optimization preference between cost, latency and quality. Name your config, set a price cap, allow or deny different models or providers and optimize between your preferences. Save it once as a reusable config tied to a specific workflow, like a fast in-app preview versus a paid final export.
Call one endpoint. Send your generation request to the Model Router endpoint with your config ID attached. No need to specify a model.
Get the best model. The router filters out any model that fails your hard constraints (capability fit, allow and deny lists, price cap), then scores the remaining models against your cost, quality and latency preferences. It generates with the highest-scoring model and returns the result along with metadata identifying exactly which model was used and why.
The initial config dimensions that will ship today are:
- Price cap: a hard ceiling per generation, in USD per second for video or per image for image models
- Allow and deny lists: provider and model-level rules that restrict routing to approved options
- Set preferences: define your preferences across cost, quality and latency optimization
If a request can't be satisfied by any model under your constraints, Model Router returns an explicit error naming which constraint emptied the pool, rather than silently downgrading your settings. And before you put a config into production, a dry-run flag lets you validate exactly which model it would select and why, with no cost incurred and nothing generated.
Model Router Scales With Your Team
Scale changes what a model choice actually costs: in dollars, in latency and in oversight.
For teams running high request volumes, small differences add up fast. A fraction of a cent or a few milliseconds of latency, multiplied across a million requests a year, becomes a material cost. Model Router also solves a real integration problem: keeping up with a fast-moving model catalog without constantly updating your code.
Try Model Router Today
Model Router is live now for video, image and audio models in Runway Dev. Start building at dev.runwayml.com.

