Runway has introduced Solaris, an AI system the company calls an "Interface World Model" — software that generates working, interactive user interfaces frame by frame in real time, without writing or executing any traditional code. The research preview, unveiled this week on Runway's site and detailed in a technical write-up, reimagines how software interfaces might be produced: not as static screens built from pre-written components, but as a continuous video-like stream generated on the fly in response to what a user does.

Solaris is built on top of Runway's Gen-4.5 video generation model and extends research first introduced with GWM-1, the company's earlier general world model. In the new system, a language model interprets what a user is asking for — a click, a drag, a typed request like "move the table" or "change the couch color" — while the world model renders the resulting visual outcome frame by frame. There is no intermediate step where a design gets translated into HTML, CSS or application code; the frame itself is the interface.

"Solaris starts with a question: what happens when an operating system generates apps and websites as you use them?" the company said in its announcement. Runway frames the project as an attempt to replace "scripted pages" with something closer to "a living environment" that responds to context rather than executing a fixed script.

Practically, the system works by starting from a seed visual state — a brand environment or a product scene, for instance — and then streaming new frames as the user interacts with it. Because it is generating pixels rather than calling predefined UI states, Runway says Solaris can support interactions that were never explicitly programmed, adapting a scene differently depending on what has already happened in it. The company points to several early use cases: adaptive e-commerce storefronts that personalize layout while keeping brand identity intact, virtual try-on tools where a user can drag a garment onto a photo of themselves, voice- or text-controlled product visualization, and interactive tutorials or educational demonstrations that render each next step contextually rather than from a fixed script.

Runway also published evaluation data alongside the announcement. In a reconstruction-fidelity test, the company had several leading multimodal models — Anthropic's Claude Fable 5, OpenAI's GPT-4o and Google's Gemini 2.5 Pro — attempt to recreate interfaces from screenshots, and reported that "every language model loses information during reconstruction," especially with complex or naturalistic imagery, a result Runway uses to motivate the case for a purpose-built interface model rather than a general-purpose one. Separately, the company ran a user study with 250 participants producing roughly 7,500 judgments comparing Solaris-generated interfaces to conventionally coded ones. Runway reports that testers preferred Solaris 61% of the time on following requested interactions and 71% of the time on natural scene behavior, versus 24% and 21% respectively for coded interfaces, with Solaris also scoring better on coherence within a scene rather than jumping between isolated UI states.

The company is upfront about where the technology still falls short. Legible, stable text remains, in Runway's own words, "extremely difficult" for the model to render reliably, which is a significant limitation for an interface-generation product given how much of ordinary software design depends on text. Runway suggests hybrid approaches — pausing generation for text-heavy views — as a practical workaround for now. The system also currently stays closely anchored to the starting imagery it's given, and Runway describes incorporating verified outside context, such as real product data or reference images, as ongoing research rather than a solved problem. Maintaining coherence over long, open-ended sessions is also unresolved, as is support for screen readers and other assistive technology, meaning Solaris in its current form has no accessibility story. Generating every frame is also markedly more computationally expensive than serving a conventional pre-built page, though Runway says that cost gap is narrowing.

Solaris is not broadly available. Runway is running an early-access program — companies can apply through a form requesting business email, company name and size — while the firm says it is "looking for partners" ahead of a wider public launch. No pricing has been disclosed. That positions Solaris closer to a research demonstration and enterprise pilot than a shipping consumer product, a distinction that matters given how easily "generates working software with no code" headlines can outrun what a system can currently deliver in production.

The announcement lands amid a broader push across the AI industry toward models that generate interactive, dynamic content rather than static outputs — a lineage that includes so-called "world models" built by labs such as Google DeepMind and various robotics and simulation research groups, most of which have targeted game environments, video, or robot training data rather than production software interfaces. Runway's pitch is distinctive in aiming Solaris directly at commercial software and web experiences, and in explicitly framing the technology as a training ground for AI agents that otherwise struggle to navigate real-world websites with constantly shifting layouts — an acknowledgment that autonomous agents built to act on the web remain a live bottleneck for the industry.

Whether Solaris becomes more than a striking research preview will depend on issues that are easy to state and hard to solve: rendering trustworthy, factually grounded content rather than plausible-looking hallucinations; making generated interfaces accessible; and bringing the cost of frame-by-frame generation down to something competitive with conventional web serving. Runway's own limitations section reads as a fairly candid roadmap of what still needs to be built before Solaris moves from early-access pilots with select partners to a general release. For now, the company says a broader public launch is being prepared "with key partners," without committing to a timeline.