Most AI demos show you a chatbot filling out a form or clicking through a website. Runway is asking a different question: what if the website itself was generated on the fly, frame by frame, as you use it? That’s the premise behind Solaris, which Runway announced on August 31, 2026 as the first model in what it calls its Interface World Model family.
This is not a UI generation tool that spits out a Figma mockup. Solaris renders an interactive interface in real time, producing each frame as you interact with it. There’s no HTML, no CSS, no JavaScript running underneath. The model is the application.
What Solaris actually does
Every piece of software built today requires a translation step. A designer produces a visual, and then an engineer converts that into code so it can actually respond to user input. That conversion is lossy by definition. You lose visual fidelity, and you lock in every possible interaction at build time. If a user tries something the developer didn’t anticipate, nothing happens.
Solaris removes that step. It generates the visual and the interactive response from a single model, with each frame conditioned on what just happened. Click something, drag something, type something, and the model synthesizes the next frame based on that input. No predefined workflow, no scripted states.
Runway describes three properties that follow from this architecture:
- Visual: The image is the application. No second implementation layer exists beneath what you see. Runway demos this with a virtual clothing store where you drag items onto a photo of yourself to try them on.
- Alive: Because frames are continuously rendered, the interface doesn’t wait for discrete user actions. Lighting shifts, objects respond physically, and users can give natural language instructions like “move the table” or “change the couch color.”
- Open-ended: Traditional interfaces only support interactions developers planned for. Solaris can respond to behaviors that were never explicitly defined, because the model determines what’s possible, not a fixed codebase.
The technical problems Runway had to solve
Runway is transparent about why this hasn’t existed before, and the engineering challenges are real. Interactivity breaks down around 500 milliseconds of delay. Standard video diffusion models take seconds or minutes per clip. So Solaris generates frames sequentially, each one depending only on prior frames, fast enough to stay ahead of user input.
Coherence is the second problem. Generated video has historically struggled to maintain consistent text, layout, and object identity across a long session. Small errors compound. Runway built Solaris to hold coherence across an entire session, not just a short clip. And it targets 720p visual quality, which is high enough to matter for real interfaces.
Cost is the third issue. Rendering every frame is more expensive than serving a static page built once. Runway says Solaris is orders of magnitude cheaper to run than a standard video diffusion model, and it’s built on top of Gen-4.5, the same video generation model that powers its existing tools.
Why this matters for AI agents
There’s a second use case here that’s arguably more interesting than the consumer interface story. Today’s LLMs are notoriously bad at computer use tasks. Booking a hotel, ordering groceries, navigating an unfamiliar website, these are tasks that trip up even the best models. Part of the reason is that agents train on specific interfaces and fail to generalize when the layout changes even slightly.
Solaris could be a training environment for agents where the interface is never the same twice. Because the model generates the interface dynamically, agents can practice on layouts that have never existed before. That’s a fundamentally different approach to agent training than scraping screenshots of real websites.
What’s missing and what to watch
Runway hasn’t released Solaris publicly yet, and pricing and availability details are not disclosed. This is a research introduction, not a product launch. There’s no API, no waitlist, and no clear timeline.
The competitive context matters here. Google, Microsoft, and several startups are all working on computer use and agent infrastructure. But most of that work assumes static, coded interfaces as the environment. Runway is betting that the environment itself should be generative. Whether that bet pays off depends on how fast the cost curve improves and whether the coherence holds up in real-world sessions, not just curated demos.
Still, the underlying idea is hard to dismiss. If interfaces can be generated rather than coded, the entire assumption that software must be built before it can be used starts to look shaky. That’s worth paying attention to.
