Google had held the top spot in the text-to-video leaderboards with its Veo 3 model for a while, but it’s now been upstaged by a plucky startup.
Runway, the New York-based AI company that pioneered publicly available video generation two years ago, has reclaimed its position at the forefront of generative video with the launch of Gen-4.5. The new model scored 1,247 Elo points on the Artificial Analysis Text to Video leaderboard, surpassing Google’s Veo 3 (1,226 points) and every other AI video model currently available.
The achievement marks a significant milestone in the rapidly evolving text-to-video space, where competition has intensified dramatically over the past year. While tech giants like Google, Meta, and OpenAI have thrown substantial resources at video generation, Runway’s latest release demonstrates that focused innovation from a specialized startup can still outpace the industry’s largest players.
From Whisper Thunder to Industry Leader
Gen-4.5, previously code-named “Whisper Thunder” on blind test platforms and internally referred to as “David,” represents what Runway calls a fundamental advancement in both pre-training data efficiency and post-training techniques for video models. The company positions it as their new foundation model for world modeling, a term that refers to AI systems capable of understanding and simulating physical reality.
The model’s capabilities extend well beyond simple text-to-video conversion. Gen-4.5 excels at understanding and executing complex, sequenced instructions, allowing creators to specify detailed camera choreography, intricate scene compositions, precise timing of events, and subtle atmospheric changes all within a single prompt. This level of control addresses one of the persistent challenges in generative video: translating creative vision into executable AI instructions.
According to Runway, the model achieves unprecedented physical accuracy and visual precision. Objects move with realistic weight and momentum, and surfaces behave as they would in the real world. Critically, the system can both observe and intentionally violate physics, depending on the creator’s vision, offering flexibility for both realistic and stylized content.
Technical Foundation and Limitations
Runway developed Gen-4.5 entirely on NVIDIA GPUs, spanning initial research and development, pre-training, post-training, and inference. The company collaborated extensively with NVIDIA to optimize video diffusion models, from training efficiency to inference speed. Model inference runs on NVIDIA’s Hopper and Blackwell series GPUs, delivering what Runway describes as an industry-first approach that balances performance with quality.
Despite the substantial leap in capabilities, Runway acknowledges that limitations remain. The company specifically cites failures in causal reasoning and object permanence as areas requiring further development. These shortcomings, common across current video generation models, represent fundamental challenges in teaching AI systems to understand how the physical world works over time.
Runway’s Competitive Position
The launch of Gen-4.5 reinforces Runway’s position as a serious competitor in the generative video market. Founded in 2018 by Cristóbal Valenzuela (CEO), Alejandro Matamala (chief design officer), and Anastasis Germanidis (CTO) at NYU, the company has evolved from a model directory for machine learning deployment into a comprehensive creative suite.
Runway reported $121.6 million in revenue for 2024 from approximately 100,000 customers, with annual recurring revenue doubling from $48.7 million in 2023. The company raised $50 million in Series C funding at a $500 million valuation, reflecting investor confidence in the commercial viability of generative video tools.
The company’s customer base spans advertising agencies, filmmakers, and major brands including New Balance and Under Armour. Runway’s tools have been used in production workflows for programs like The Late Show With Stephen Colbert and in Oscar-nominated projects, demonstrating real-world adoption beyond experimental use cases.
The Broader Text-to-Video Landscape
The Artificial Analysis leaderboard reveals a highly competitive field. Following Runway Gen-4.5 and Google’s Veo 3, Kuaishou’s Kling 2.5 Turbo (1,225 points), Google’s Veo 3.1 variants, Luma Labs’ Ray 3 (1,211 points), and OpenAI’s Sora 2 Pro (1,206 points) round out the top performers. The clustering of scores suggests that multiple approaches to video generation are converging on similar capability levels.
What differentiates these models increasingly lies not just in raw quality scores but in specific strengths: prompt adherence, motion quality, visual fidelity, controllability, and generation speed. As the technology matures, use-case-specific optimization may become more important than general-purpose supremacy.
Rollout and Future Development
Runway is gradually rolling out access to Gen-4.5, with availability expanding to all users in the coming days. The company plans to bring all existing control modes to the new model, including Image to Video, Keyframes, and Video to Video, alongside additional capabilities.
The launch represents more than just a leaderboard victory. It signals that the text-to-video generation space remains open to innovation from companies of all sizes, and that the technological moat around generative AI may be narrower than the resource advantages of major tech companies would suggest. For Runway, maintaining this lead will require continued innovation as competitors inevitably respond with their own advances. And for Google, it shows that while it has many impressive strides in AI over the last year, it still has smaller startups nipping at its heels.