Runway launched Gen 4.5 yesterday, claiming the top spot on the Video Arena leaderboard with an Elo score of 1,247. The text-to-video AI model from the 100-person New York startup displaced Google’s Veo 3 and pushed OpenAI’s Sora 2 Pro down to 7th place. It’s a rare David vs Goliath win in an AI arms race dominated by trillion-dollar tech giants.
The rankings aren’t based on marketing claims—Video Arena uses blind comparisons where users vote between two model outputs without knowing which company produced them. Runway beat both Google and OpenAI on the metrics that matter most: prompt adherence and motion quality.
Specialization Beats Scale
Runway CEO Cristóbal Valenzuela put it bluntly: “We managed to out-compete trillion-dollar companies with a team of 100 people.” The startup has focused exclusively on video generation since founding in 2018, while Google and OpenAI spread resources across language models, code generation, robotics, and multimodal AI.
That specialization paid off. Gen 4.5 excels at physics simulation—objects move with realistic weight and momentum, liquids flow with proper dynamics, and fine details like hair strands stay coherent across motion. These aren’t incremental improvements. They’re the result of seven years spent solving one problem: making AI-generated video look and feel real.
Google’s Veo 3 supports 4K video and synchronized audio. OpenAI’s Sora 2 generates longer clips with better character consistency. However, neither matched Gen 4.5 on the benchmark that measures actual video quality. Broad platforms lost to a focused tool.
Related: Databricks Raises $5B, But AI Products Squeeze Margins
The NVIDIA Advantage
Runway built Gen 4.5 entirely on NVIDIA GPUs, with inference running on the company’s cutting-edge Blackwell architecture. This is one of the first production deployments of Blackwell chips, giving Runway access to hardware most competitors can’t get yet.
NVIDIA CEO Jensen Huang endorsed the release: “This is an incredibly exciting time for video and world models.” Moreover, NVIDIA has invested in Runway through multiple funding rounds and clearly sees the startup as a strategic partner. Access to next-gen chips before they’re widely available is a significant moat.
December 1: Super Sunday for AI Video
Runway wasn’t the only company launching yesterday. ByteDance released Vidi2, and Tencent shipped HunyuanVideo-1.5—three major AI video models on the same day. The coordinated timing suggests companies are racing to stake claims before the market consolidates.
The competition is intensifying fast. Furthermore, Google, OpenAI, Meta, ByteDance, and now Tencent are all pushing video generation models. Runway’s #1 position today doesn’t guarantee dominance tomorrow. Consequently, leaderboard rankings shift quickly as models update.
Available This Week at Same Pricing
Gen 4.5 is rolling out gradually to all Runway users within days. Additionally, pricing matches Gen-4: Standard ($15/month, 625 credits), Pro ($35/month, 2,250 credits), Unlimited ($95/month). Video generation costs 12 credits per second for standard quality, 5 credits per second for Turbo mode.
No price increase despite performance gains. In fact, Runway is prioritizing market share over margin expansion, which makes sense given how crowded the space is getting. API access is available for developers who want to integrate Gen 4.5 into custom workflows.
Limitations Still Exist
Runway disclosed Gen 4.5’s current limitations in their research blog: causal reasoning issues (effects sometimes precede causes), object permanence problems (things disappear or appear unexpectedly), and success bias (actions succeed more often than they would in reality).
These are common issues across all AI video models, not unique to Runway. Nevertheless, it’s worth noting: Gen 4.5 is the best available, not perfect. Creators still need to review outputs carefully and be prepared to regenerate or manually fix issues. The technology is production-ready for many use cases, but it’s not replacing human judgment yet.
Key Takeaways
- Runway Gen 4.5 tops Video Arena leaderboard (Elo 1,247), beating Google Veo 3 (#2) and OpenAI Sora 2 Pro (#7)
- Specialization advantage: 100-person team focused on video beats tech giants spreading resources
- NVIDIA Blackwell deployment gives Runway infrastructure edge over competitors
- December 1 “Super Sunday” saw three major AI video releases (Runway, ByteDance, Tencent)
- Available this week at same pricing as Gen-4 (no premium despite performance gains)
- Limitations persist: causal reasoning, object permanence, success bias still need work










