Runway Solaris: The Interface World Model, Explained

Runway is best known for AI video, but its latest release points somewhere stranger: software that is generated, not written. On August 31, 2026, the company introduced Solaris, which it calls the first model in a new family of Interface World Models — systems that render interactive user interfaces in real time, frame by frame, with no underlying code. Click a button and there is no event handler firing; instead, the model paints the next frame showing what the button did. The interface is not running. It is being imagined, continuously, as you use it.

The idea sits at the intersection of two threads Runway has been pulling on for a while: generative video and world models, the AI systems that learn to simulate an environment and predict how it evolves under a user’s actions. Solaris applies that machinery to a new domain — the graphical interface itself — and the result is a research demo that reframes what “an app” could be.

How Solaris works

Conventional software separates state from rendering. A program holds data, logic decides what should change when you interact, and a rendering layer draws the result. Solaris collapses that stack. A single world model generates every frame and every response to user input, with no intermediate representation — no DOM, no component tree, no code the model is executing behind the scenes.

The interaction loop works like this: your clicks, drags and typed instructions become inputs to the next frame. A language model interprets what should happen — what a reasonable interface would do in response — and the world model paints that decision directly onto the screen. Type “add a row to the table” and the model renders a table with a new row; drag a slider and it renders the frames of the slider moving and whatever the interface would show as a consequence.

Technically, Solaris builds on Runway’s Gen-4.5 video model and follows the company’s broader GWM-1 general-world-model research. The key architectural choice is that GWM-1-style models are autoregressive in time: rather than generating an entire clip at once the way a standard diffusion video model does, they produce one frame at a time, conditioned on past frames and the latest control input. That is what makes real-time interactivity possible — the model only has to imagine the next moment, given what you just did, instead of re-rendering a whole timeline. Runway has described this as the same path it opened with GWM-1, now specialized for interfaces.

The benchmark claims

Runway says Solaris outperforms frontier large language models at generating new interfaces, specifically on measures of structural similarity and information retention — roughly, whether the generated interface looks like a coherent, correct layout and whether it keeps track of state as you interact with it.

The company also ran a human-preference study, and the setup is worth understanding because it frames Solaris against a specific rival. Runway compared Solaris with a coded interface generated by Claude Opus 5, with both systems starting from the same image and receiving the same interaction instructions — one producing frames, the other producing code that renders an interface. Across a study of about 250 people, 30 examples and nearly 7,500 pairwise judgments, participants preferred Solaris 61% to 24% for following the requested interaction, and 71% to 21% for behaving naturally within the scene.

Those are strong numbers, but they come with an important asterisk: they are Runway’s own results, and at announcement there were no independent benchmarks and no third-party verification of the claim that Solaris surpasses coding models at interface generation. Human-preference studies are also sensitive to how examples are chosen and how “natural” is defined, so the figures are best read as a promising internal signal rather than a settled comparison. The contrast with a code-generating model like the current Claude model lineup is the whole point of the pitch — and the whole thing third parties will want to test.

What it is not (yet)

Solaris is a research release, not a product. As of September 1, 2026, Runway had announced the model, opened an early-access request, and published no public pricing, no API and no general sign-up. There is no way for a developer to build on it today, and the demos on offer are curated. That places Solaris firmly in the “look what’s possible” category rather than the “ship this in your app” one.

There are also open questions the demo does not answer. Frame-by-frame generation of an interface is expensive — every interaction is an inference step — which raises real cost and latency concerns for anything beyond short, contained sessions. Determinism is another: traditional software does the same thing every time you click the same button, while a generative interface can, in principle, improvise, which is delightful in a demo and alarming in, say, a banking app. And nothing generated as video frames is directly inspectable, testable or version-controlled the way code is.

Where it fits in the world-model wave

Solaris is part of a broader surge of investment in world models as a distinct branch of AI, separate from the language models that have dominated the last few years. The bet across the field is that models which learn to simulate environments — physics, robotics, game worlds, and now interfaces — unlock capabilities that pure text prediction cannot. That thesis has been drawing serious money; the recent $90 million seed round for a world-models startup led by Sanja Fidler is one marker of how much investor conviction has moved toward simulation-first systems.

It also lands in a market already being reshaped by AI-assisted software creation. The dominant paradigm today is code generation — models that write the HTML, CSS and application logic that a browser then runs, a shift documented in the state of AI coding assistants in 2026. Solaris proposes something categorically different: skip the code entirely and generate the running experience. Whether that is a better path or a detour is the argument Runway is starting.

What it means

Solaris is a genuine research milestone and a legitimately new idea about what an interface can be — but it is a long way from changing how software gets built. The honest read is that Runway has shown a striking proof of concept and made bold, self-reported claims about beating code-generating models, without yet giving anyone outside the company a way to verify them or use the thing.

Who benefits. Runway extends its world-model research from video into a domain — interfaces — that touches every piece of software, strengthening its position as a serious lab rather than a single-product company. Researchers gain a concrete new testbed for interactive, autoregressive generation. And designers and prototypers stand to benefit first, since imagining an interface from an image and a few instructions is exactly the kind of fast, throwaway exploration a generative model is well suited to.

Who should stay skeptical. Engineers building production software have little to act on here: cost, latency, determinism, testability and inspectability are all unsolved, and none of them are minor. The head-to-head against a coding model is Runway’s own study, and “beats the code model” is a claim that needs independent replication before it means anything for real workflows.

What to watch next. The signals that would move Solaris from demo to platform are concrete: an API or public access that lets outsiders reproduce the results; independent benchmarks on interface generation; and evidence that frame-by-frame generation can be made cheap and consistent enough for sessions longer than a demo. If Runway delivers those, the “generated, not coded” interface becomes a real design space. If it does not, Solaris will still matter — as the moment someone showed the alternative was even possible.