Runway released Solaris, a model that generates app and website interfaces frame by frame as a person interacts with them, rather than running pre-built code. The company calls it the first of a new category, an Interface World Model, and is opening early access through a signup form rather than a public launch. The pitch inverts a design assumption that has held since the first graphical operating systems: that a visual mockup is inert until somebody compiles it.
That compilation step is the target. Runway argues every interface built today loses information twice: once when a design gets reduced to buttons, layouts and event handlers, and again when a developer has to predict every interaction a user might attempt. Solaris skips both steps by treating a click or a drag as another input the model conditions on, the same way it conditions on text or a reference image, and generating the next frame directly.
Solaris is built on Runway’s Gen-4.5 video model, adapted to generate frames autoregressively and distilled down to a handful of denoising steps so it can respond at interactive speeds rather than the seconds or minutes a standard video diffusion model needs per clip. A separate language model decides what an interaction should trigger; Solaris renders how it looks. Runway says this split, reasoning in one system and rendering in another, is what lets the same scene support behaviors nobody explicitly programmed, such as dragging a shirt from a rack onto a photo of yourself to try it on.
Runway backs the pitch with two evaluations. In the first, it tested how well multimodal models, including Claude Fable 5, could reconstruct a website from a single screenshot, and found reconstruction quality dropping as visual complexity rose, evidence for its claim that translating a design into code or language loses information. In the second, it ran a user study of 250 participants judging 30 interaction examples across roughly 7,500 pairwise comparisons, pitting Solaris against a website Claude Opus 5 coded from the same starting image. Participants preferred Solaris’s output 61% of the time for following the requested interaction and 71% of the time for feeling natural within the scene. Those are Runway’s own commissioned figures, not an independent audit, and the company has not said whether the comparison used a single prompting approach for the coded baseline or several.
Runway is candid about what Solaris cannot yet do: it struggles to render stable, legible text, has no verification layer to stop a confident but wrong answer, and has not been tested for coherence over long sessions or compatibility with screen readers and other accessibility tooling. Those are the gaps that separate a striking demo from something a business could put in front of customers.
The deeper trade is one Runway does not fully spell out. A model that generates the interface itself, rather than the code for one, inverts how software has been built and shipped for forty years, and the honest question is what that costs in determinism: whether a click that produced one result once will produce the same result the second time, and how a business audits or debugs behavior that was never written down as logic anywhere.
Runway also frames Solaris as a training environment for computer-use agents, which today largely fail outside the exact interface layouts they were trained on. If Solaris can generate interfaces that vary the way real websites do, it could become a testbed for agents that generalize past memorized layouts, a capability current benchmarks suggest is still missing. Teams building browser or computer-use agents should watch Runway’s early-access rollout for whether that training use case ships before, or instead of, the consumer interface product.
Runway detailed Solaris in a research post on its own site introducing the model, titled “Introducing Solaris.”
