On August 31, Runway introduced Solaris, the first release in a category it is calling Interface World Models. The pitch: a user clicks, drags, or types, and the model generates the next frame of the interface directly, without a design file, a component library, or a line of JavaScript in between. Runway is calling this the end of the design-to-code pipeline. That claim is worth taking seriously and worth testing against what Runway has actually shipped, which as of this week is a research demo, not a product.

What Solaris actually does

Solaris is built on Runway's Gen-4.5 video model and extends the approach behind GWM-1, Runway's general world model released in December 2025 for 3D environments and avatar training data. Where GWM-1 modeled physical scenes, Solaris models software. It observes clicks and drags as inputs and predicts, frame by frame, what the screen should look like next. There is no DOM to inspect, no CSS to override, no event handler firing behind the button. The button is a pixel pattern, and the model paints what happens after it is pressed.

Runway says Solaris outperforms frontier language models on two specific measures: structural similarity, meaning whether the generated layout looks coherent, and information retention, meaning whether the interface remembers state correctly as a user keeps interacting with it. Those are the right benchmarks to watch, and they are also the two places generative interfaces have historically failed, so treat early numbers as directional until independent testing catches up.

The moat Runway is chasing

If an interface can be generated rather than assembled, the competitive advantage stops living in who has the best design system or the fastest release cycle. It moves to whoever holds the largest library of real interaction data, the clicks, drags, and corrections that teach a model what a good interface does next. That is a genuine strategic bet, and it explains why Runway is building this on top of a video model rather than a language model. Video generation already had to solve temporal coherence. Interfaces just add a harder constraint: the state has to be correct, not merely plausible.

The technique itself is not exclusive to Runway. Autoregressive frame generation conditioned on user input has shown up in other 2026 research, including work on causal forcing for video models. What Runway is betting on is not a secret architecture. It is being first to point the approach at software interfaces at product scale, and to own the interaction data that comes from being first.

What's missing before this matters to your budget

As of this week, Solaris has an early-access request form. It has no public pricing, no API, and no general release. That matters more than it sounds like it should. Every frame of every interaction now has to be generated on a GPU instead of served from a cache, which is a fundamentally different cost curve than hosting static assets on a content delivery network. Runway has not said what that curve looks like at scale, and until it does, any estimate of frontend budget savings is a guess dressed up as a forecast.

None of this is a reason to dismiss the category. Interface World Models are the logical next step after a year of world models for physical scenes, and Runway shipping two of them nine months apart says something about how fast this research is moving. It is a reason to separate the demo from the deployment timeline before it shows up in a strategy deck.

The frontend stack has survived every framework war of the last decade. What it has not yet faced is a competitor that skips the framework entirely.

Watch three things over the next two quarters: whether Runway publishes an inference cost per session, whether early-access partners include an enterprise SaaS vendor rather than only consumer app builders, and whether a second lab announces a competing Interface World Model. Any one of those would move this from research curiosity to procurement conversation.