Runway's Gen-4.5 tops the video arena leaderboard, beating Google and OpenAI with a research-first, taste-driven approach

Summary

Runway's Gen 4.5 has taken the top spot on the Video Arena Leaderboard, a crowdsourced ranking that pits anonymous model outputs head-to-head for public votes, displacing both Google and OpenAI. Cristóbal Valenzuela, Runway's CEO, frames the result as a validation of research discipline over raw capital, citing Ilya Sutskever's observation that the industry has re-entered an era of research as a tailwind.

The competitive edge, per Valenzuela, is not compute scale but what he calls "taste" — a deliberate, top-down curatorial philosophy baked into training from the start. Where Sora's outputs read as trained on short-form social content and Veo 3 skews toward high-saturation cinematic excess, Runway targeted professional-grade, controllable output for filmmakers and creative studios. That distinction is intentional: training choices reflect business model choices, and Runway is explicitly building for creatives, not advertisers.

On capital efficiency, Valenzuela stops short of specifics but is clear that Gen 4.5 was not built on a tens-of-billions-of-dollars training budget. The more durable asset, in his framing, is the institutional knowledge Runway has accumulated across nearly a decade of model development — something that cannot be replicated by spending alone. Inference pricing for Gen 4.5 is described as comparable to prior Runway models, with near-real-time generation on the roadmap.

On Hollywood adoption, Valenzuela pushes back on the narrative of industry resistance, arguing it is a vocal but unrepresentative minority. Studios, agencies, and production teams are already using AI tools, in his telling, and the public discourse has simply not caught up. He positions gaming companies as the sector now entering the adoption curve that Hollywood crossed roughly 18 months ago.

On the NLE question — whether an AI-native nonlinear editor is the next platform battleground — Valenzuela is skeptical of the premise. Average scene duration in film already runs two to three seconds, meaning long-sequence generation is a less important problem than multi-shot narrative coherence. More broadly, he sees the NLE paradigm itself as transitional, likely to be rendered obsolete as real-time AI generation removes the need for asynchronous editing workflows. Current integrations of AI into tools like Premiere Pro or DaVinci are, in his view, a bridge, not a destination.

Runway's unlimited plan offers 2,250 credits monthly with queue-based generation, and Valenzuela claims no competitor currently offers an equivalent unlimited tier.