Runway says its model router cuts video costs
Runway says a $1-capped configuration of its Model Router cut video-generation spend by 66% against a Seedance 2.5 baseline while keeping 95% of its usable-output rate, according to the company's own benchmark.
Runway has published a benchmark of Model Router, a Runway Dev feature that chooses a generative media model for each request. Builders set a preference for cost, quality or latency, and the router uses Runway's internal benchmarking data to pick the model.
The test ran 250 image-to-video prompts, split evenly across 10 categories, including physics, text and signage, lip-sync and product shots. Runway compared three router configurations against a fixed baseline of Seedance 2.5:
| Configuration | Usable rate | Cost/clip | Spend vs. baseline | |---|---|---|---| | Seedance 2.5 baseline | 78% | $1.80 | — | | Quality router | 77% | $1.28 | -29% | | Quality + $1 cap | 74% | $0.61 | -66% | | Cost router | 38% | $0.30 | -83% |
The uncapped quality router still sent 56% of prompts to Seedance 2.5. Most of the rest went to Gemini Omni Flash 1.1 and 1.0. With the $1 cap, all calls went to Gemini Omni Flash 1.1 (65%), Gemini Omni Flash 1.0 (34%) and Happy Horse (1%). The cost router sent 99% of requests to Runway's Gen-4 Turbo, and its usable rate fell by half. Runway suggests that setting fits drafting work, with final outputs sent through a quality router.
Runway says the evaluation was double-blinded and used paired two-sided t-tests with Benjamini-Hochberg correction. It does not explain how "usable" was judged or who did the judging. The company also reports that 64% of Model Router users have set their routers to optimize for cost.
Runway designed and ran the test on its own product, and no independent party has checked the results.