Disclosure: TechSifted has no affiliate relationship with Runway. This is editorial news coverage of a publicly announced research preview.
Runway is not a video generation company anymore. Or at least, that's what Runway is trying to tell you.
On August 31, 2026, Runway announced Solaris — a research preview they're calling the first Interface World Model. It doesn't generate video clips you export and use somewhere else. It generates a live, interactive software interface, frame by frame, with no code underneath. You interact with the screen. The model renders the next frame based on what you did. There's no HTML. There's no CSS. There's no DOM. There is no software in the traditional sense — just a model generating pixels in real time that behave like software.
That's the bet. And it's a large one.
What Solaris Actually Is
Let me be clear about the mechanics before we talk about what they mean, because "AI generates software interfaces" is the kind of headline that gets distorted in both directions — either inflated into a revolutionary demo or dismissed as vaporware.
Solaris is built on Runway's Gen-4.5 video model, the same model that powers Runway's video generation tools. The key architectural insight is that a video model generates frames sequentially, where each frame depends on what came before. In video generation, those prior frames are previous images in a clip. In Solaris, they're previous states of an interface — and user interactions like clicks and drags are treated as conditioning signals for the next frame, the same way a text prompt or reference image is used in video generation.
The result: a live 720p interface that responds to your actions, without any underlying DOM or accessibility APIs.
Runway splits the intelligence across two systems. A language model handles the reasoning layer — it interprets what you're asking, decides what should change, and generates prompts. A world model handles the rendering layer — it synthesizes what the resulting state looks like. One brain, one paint brush, passing instructions back and forth at interactive speeds.
In Runway's own user study — their numbers, their study — people preferred Solaris over a coded interface built by Claude Opus 5 by 61% to 24% on following instructions, and 71% to 21% on behaving naturally. Those are striking numbers from a research preview. They're also Runway's numbers, so read them with the skepticism that requires.
The Strategic Move Here
There's a pattern in how AI companies evolve: you start with the most legible, demonstrable version of your capability, prove the model can do something impressive, and then use that foothold to expand the surface area of what the model is for.
OpenAI started with text generation, added images, then video, then voice, then agents. Google did the same. Runway started with video generation, built the best cinematic text-to-video model on the market, accumulated the training data and engineering infrastructure that comes with that, and is now using the underlying world-model architecture to build something that generates interactive software rather than passive clips.
The pivot from video to interfaces is more natural than it sounds. Runway's Gen-4 lineage wasn't just trained to generate pretty clips — it was trained to model how things move, how visual states transition, how the world looks in response to change. Solaris applies that same capability to a different domain: instead of modeling how a scene looks after an action in the world, it models how a screen looks after a user action. The architecture is different enough to be a genuine research contribution. It's similar enough to what Runway already does that building it made sense for them and probably no one else.
This is also a land grab. The Rundown AI framed Solaris as previewing "the no-code internet," and that framing captures the competitive logic: if an interface can be generated rather than assembled, the value in software shifts.
That's a very different business than selling subscriptions to a video editor.
Read the rest of this article on TechSifted →
Originally published at techsifted.com.
Top comments (0)