On August 31, Runway introduced Solaris, the first in a new category it calls Interface World Models. Instead of writing code that a browser renders, Solaris generates each screen as a video frame, in real time, as you interact.
How it works
Solaris pairs Runway's Gen-4.5 video model with a language model. When you click or drag, the LLM decides what should happen next, then the video model renders that decision as the next frame. Demos include dragging a shirt onto a model, building a salad by dropping ingredients, and a working combustion simulation. There is no HTML, CSS, or JavaScript anywhere in the pipeline.
Runway says a 250-person user study preferred Solaris over interfaces coded by Claude Opus 5 in 61% of instruction-following tests and 71% of natural-behavior comparisons. The company also measured higher structural similarity and information retention than frontier LLMs on interface reconstruction tasks.
What still breaks
Runway's own paper flags text legibility, session drift over long interactions, and screens that look correct but behave wrong. Accessibility tools that rely on reading the DOM—the code structure behind a webpage—have nothing to read here. The interface is just pixels. Solaris is entering early access by request, with no public pricing yet.
The tension nobody mentions
If Solaris works at scale, you stop needing to own or understand code to ship software. That sounds like democratization. It also means the model provider becomes the gatekeeper to every button, menu, and interaction you build. You are not writing the rules anymore—you are renting the physics.