The interface is no longer a design file. It is a video being generated.
On August 31, Runway released Solaris. It is not "another AI that writes front-end code for you." Runway calls it the first of a new family called Interface World Models. The difference matters. An AI that writes front-end code outputs code, and the browser renders that code into a page. Solaris outputs the picture itself. There is no DOM underneath, no CSS, no JavaScript.
You click, you drag, and the model generates the next frame. In a clothing store, you drag a shirt onto a photo of yourself and the shirt goes on. In a room planning app you say "move the table so I can see how it looks," and the scene responds. The whole thing feels like playing a video, except the video is being made as you interact with it.
How it works
The base is Runway's own Gen-4.5 video model, following the technical path of its earlier GWM-1 world model. The architecture splits the work in two. A language model figures out what you want and decides how the scene should change, then writes a prompt. The world model takes that prompt, renders the change as a picture, and keeps producing the frames that follow.
Three changes compressed a process that used to take seconds or minutes into something real time: generating frames autoregressively one at a time, distilling away most of the denoising steps, and training the fast model on its own outputs so it stays coherent over long sessions. Runway's published figure is roughly 37 milliseconds per frame at 720p, which works out to about 27 frames a second. That reads as continuous motion to the eye.
The numbers are real. Do not celebrate yet.
Runway ran its own study: 250 participants, 30 interaction examples, close to 7,500 pairwise judgments. On instruction following, Solaris was preferred 61% of the time against 24% for coded interfaces, with 13% judged equivalent. On natural behavior, it was 71% against 21%.
That is a wide gap. Two caveats travel with it. First, this is Runway's own study, with its own participants and its own examples. Second, 61% means that in nearly four cases out of ten, people still preferred the boring interface where the buttons stay exactly where they were. Some outlets reported that Solaris beats GPT-4o, Gemini 2.5 Pro and Claude Opus 5 on interface reconstruction, but there is no independent third-party verification of that claim. Treat it as a rumor for now.
Three problems that are not solved
Runway is candid about these.
Text is the first one. Interfaces are mostly made of text, and Solaris is a video model. Producing stable, legible, readable text is precisely the hardest part of video generation. If the label on a button comes out slightly warped or blurry, the interface stops working.
Long-session consistency is the second. Run it long enough and the picture drifts. The shirt you dragged over three minutes ago may come back a different color. Runway says the self-training step exists for exactly this problem, but no long-duration measurement has been published.
Accessibility is the third. Screen readers, accessibility APIs, that entire software stack: Solaris does not connect to it yet. For something positioned as a replacement for interfaces, that is close to disqualifying, because no real shipped product gets to skip it.
There is a quieter one underneath: a generated interface can look completely convincing while showing wrong information. Runway says it is working on keeping Solaris grounded in verified product information and reference material. In plain language, that grounding does not exist yet.
So what is it good for right now
Solaris is a research system. It has not been released publicly. Runway is taking early-access requests and working with a small set of partners toward a launch. Nobody should be putting this in production this year.
Two uses look genuinely promising. One is rapid prototyping: between an idea and a clickable, draggable demo there is no longer a design file and a front-end engineer. Changing direction is a sentence. The other is more interesting. Solaris can serve as a training environment for computer-use agents. Today's agents memorize button positions on fixed pages and fall apart on any layout they have not seen. A system that can endlessly generate novel interfaces gives them somewhere to learn generalization.
What this means for you
If you design or build front ends, do not panic, and do not look away either. Panic is unwarranted, because the things a production interface must guarantee are exactly the things a generated one cannot: that a button is the same button every time, that a transaction completes exactly once, that the thing passes accessibility review. Coded interfaces guarantee those by construction. That is a difference in approach, not a gap two more versions will close.
But the stretch from idea to something you can click and drag is about to get much shorter. It used to take mockups, asset cutting and prototype code. Now it may take a clearly stated sentence. The valuable work moves upstream: deciding what is worth building, saying precisely what you want, and defining what counts as correct.
One practical suggestion. Move your prototyping step onto tools like this and see what happens. For anything that actually ships, write the code.