Runway drops Gen-4.5 to reclaim the video crown

Runway just dropped Gen-4.5, taking the crown in AI video generation with 1,247 Elo points on the Artificial Analysis benchmark. The new release edges past Google’s Veo 3 (1,226), Kling 2.5 (1,225), and OpenAI’s Sora 2 Pro (1,206) in blind pairwise human voting.

The headline upgrade here is physical fidelity. Liquids actually behave like fluid, objects carry momentum and weight during collisions, and intricate details like hair strands stay sharp across movement. Built and optimized on NVIDIA’s Hopper and Blackwell GPUs, Gen-4.5 keeps Gen-4 speeds while significantly improving prompt adherence and aesthetic versatility, spanning photorealistic scenes to stop-motion styles.

Gen-4.5 is rolling out to all paid Runway subscription tiers at no extra charge, with existing features like image-to-video, keyframes, and video-to-video coming soon.

To their credit, Runway also published a candid list of current blind spots. The model still suffers from causal confusion (doors swinging open before handles turn), object permanence failures (items vanishing after being blocked), and a curious "success bias" where bad kicks still score goals. Runway notes that fixing these quirks is the core hurdle toward genuine world models.

Why it matters

For filmmakers, ad agencies, and indie creators, Gen-4.5 brings top-tier cinematic coherence without bumping subscription fees. Reduced visual glitching means far fewer discarded generations when building continuous camera moves.

It also heaps pressure on OpenAI and Google. With Sora 2 Pro trailing by 41 Elo points on Artificial Analysis, Runway proved that fine-tuning physical realism alongside NVIDIA hardware optimization can outpace pure hype. We think Runway's candid admission of physical reasoning flaws is refreshingly honest, but it also shows that today's best video models are still simulating appearances rather than truly understanding how our physical world works.