Runway Builds the First AI Router Spanning Video, Image, and Audio
Runway's new Media Router lets developers set quality, cost, or latency preferences once and automatically routes each request to the best video, image, or audio model.
- Runway Media Router is the first preference-optimized router for generative media, live now in Runway Dev.
- Developers set a preference (cost, latency, or quality) once; the router automatically picks the best video, image, or audio model per request.
- The router covers Runway's own models (Gen-4.5, Aleph 2.0) plus third-party models including Veo 3.1, Seedance 2, GPT Image 2, and ElevenLabs audio.
- Billing is transparent: routed generations are charged at the selected model's standard rate, with the response reporting which model ran and the exact credit cost.
- The intelligence layer is built on Runway's in-house creative team expertise -- evaluating how models handle motion, composition, and lip sync across media types.
- The launch signals Runway's broader pivot from AI model maker to generative media infrastructure layer, competing on orchestration as its own models lose leaderboard positions.
Model routing has been a staple of LLM infrastructure for a while. You set a preference, the system picks the cheapest or fastest model that meets your bar, and you stop worrying about it. Runway just brought that same idea to generative media, covering video, image, and audio generation under a single router for the first time.
A catalog too large to manage by hand
The generative media model landscape has exploded. Through Runway Dev alone, developers can already access Gen-4.5, Aleph 2.0, Seedance 2, Veo 3.1, GPT Image 2, Gemini Omni Flash, ElevenLabs audio, and more. Keeping up with which model handles motion well, which one nails composition, and which one falls apart on lip sync is a full-time job most teams can't afford. Runway's Media Router encodes that expertise so developers don't have to.
Runway says this is the first model router built specifically for generative media rather than language models. The intelligence underneath it comes from Runway's in-house creative team, who have spent considerable time evaluating outputs across every media type. With LLMs, benchmarks like MMLU give you a rough signal. With video and image models, quality is far more subjective and task-dependent, which makes that human evaluation layer genuinely useful.
How the router resolves a request
Each request goes through two steps. First, the router narrows the catalog to models that are enabled in your configuration, support the capabilities your request requires, and fall under any price cap you've set. Then it selects the best remaining model according to your stated preference: cost, latency, or quality.