Runway launches GWM-1, a real-time interactive world model

If you thought AI video companies were just here to help Hollywood save on VFX budgets, Runway has other plans.

Runway just unveiled GWM-1, its first family of General World Models built on top of its latest Gen-4.5 foundation. Instead of just spitting out pre-rendered video clips from a prompt, GWM-1 generates frames autoregressively in real time at 24 fps and 720p, reacting continuously to external inputs like camera angles, audio, and robotic controls.

Runway CTO Anastasis Germanidis explained their philosophy during the launch: teaching models to predict pixels directly at massive scale is the clearest path to simulating reality. If the model understands pixels deeply enough, it inherently grasps physics, geometry, and lighting.

Three models today, one reality tomorrow

Rather than launching a single monolithic system, Runway split GWM-1 into three specialized post-trained models, with plans to unify them later under one roof.

First up is GWM Worlds, an interactive environment engine. Feed it a text prompt or an image, and you can explore an endless virtual space in real time. Runway envisions this for open-ended gaming, but more importantly, as a spatial playground to train autonomous agents.

Then there is GWM Robotics, backed by a dedicated Python SDK. It brings diffusion-based visual representations directly into robot policy models like OpenVLA and OpenPi. Roboticists can evaluate policies in simulation, generate synthetic training data with shifting obstacles or weather, and test counterfactual trajectories without risking physical hardware.

The third leg is GWM Avatars, surfaced in the product as Runway Characters. It generates audio-driven digital humans that maintain lip-sync, subtle eye movements, and natural body gestures across extended interactive conversations.

To sweeten the release, Runway also upgraded Gen-4.5 for paid users. The flagship model now supports native audio generation, audio editing, and multi-shot consistency, allowing creators to alter an object or background in one shot and see that change propagate seamlessly across an entire one-minute video.

Why it matters

We think this is a brilliant strategic pivot that breaks Runway out of the creative tools cage.

Video generation is becoming a brutal, low-margin race against competitors like Kling, Google, and OpenAI. By repositioning Gen-4.5 as a world simulation engine, Runway is expanding its addressable market to robotics, industrial simulation, and spatial AI. Serving as the synthetic training ground for physical robots is far more lucrative than selling subscriptions to video editors.

For robotics startups, GWM Robotics could be an immediate win. Gathering real-world manipulation data is excruciatingly slow and expensive. Running policy benchmarks and safety tests inside an interactive video simulator saves both time and broken hardware.

The big question mark, however, lies in physical fidelity. Predicting pixels is not the same as computing real-world classical mechanics. An error or hallucination in a video game is funny, but in an industrial setting, a policy trained on flawed physics could wreck real machinery. Runway has built a stunning visual simulator; now it has to prove that its simulated physics can actually survive the real world.