On December 2nd Eastern Time, the video generation startup Runway announced that its latest model Gen-4.5 topped the third-party blind testing platform Video Arena, defeating Google's Veo3 and OpenAI's Sora2Pro without the participants' knowledge, becoming the first large model to achieve the top position by a small team. Runway's CEO Cristóbal Valenzuela stated that this achievement validated the feasibility of the "focus on research and rapid iteration" approach, saying, "It's not about budget, but about density, one hundred against ten thousand."
Gen-4.5 uses an in-house spatiotemporal hybrid Transformer, supporting 1080p, 24fps, and up to 18 seconds of output. It focuses on optimizing motion range, cinematography, and causal coherence. Internal testing shows that the model's success rate for complex commands such as "bullet time" and "handheld tracking" has increased to 68%, an increase of 19 percentage points from the previous generation. Valenzuela revealed that the internal code name "David" symbolizes "a small giant defeating a giant," and emphasized that the model has completed commercial closed-loop verification in several U.S. film studios.
Runway currently has 120 employees, with a valuation of $3.55 billion, backed by investors including General Atlantic, Baillie Gifford, and NVIDIA. The company plans to gradually open Gen-4.5 to global customers through its website, API, and partner channels starting this week, and will launch enterprise-level fine-tuning and on-premises deployment solutions in the first quarter of next year, continuing to expand its market share in advertising, short films, and pre-production use. Industry analysts believe that if Runway can maintain its iteration pace, the AI video sector may see a lasting competitive landscape for the first time where a small player competes with larger ones effectively.
