Runway unveils Solaris, an "Interface World Model" that generates working software frame by frame

Runway published research on August 31, 2026 introducing Solaris, which it calls the first entry in a new family of AI systems: "Interface World Models." Rather than generating code that a browser or OS then renders, Solaris generates the interface itself in real time, frame by frame, as a user interacts with it.

What's new

Runway writes: "Today, we're sharing Solaris: the first model in a new family of AI systems we call Interface World Models." The company frames the release around a specific bottleneck in how software gets built today: every interface, no matter how it's designed, has to be translated into an intermediate representation — code — before it can run, and that translation step both limits what an interface can do and strips away visual fidelity from the original design.

Solaris is built on top of Runway's Gen-4.5 video generation model, adapted to understand user interaction (clicks, drags, text input) and respond to it in real time, following on from the company's earlier general world model, GWM-1. Every user action becomes a conditioning signal for the next generated frame — the same way Runway's video models condition on text or image prompts — so the model is generating both what the screen looks like and what it does next, without an intermediate code representation.

Runway ran a head-to-head test against Claude Opus 5 generating a standard coded interface, starting both systems from the same image and interaction requests. A user study of 250 participants produced roughly 7,500 pairwise judgments across 30 interaction examples. Participants preferred Solaris's output for following the requested interaction 61% of the time versus 24% for the coded result (13% rated equivalent), and preferred it for natural behavior 71% of the time versus 21% (6% equivalent).

Context

Runway has spent the past two years pushing its video-generation models (the Gen series) toward general "world models" that simulate physics and interaction rather than just producing clips — GWM-1 was an earlier step in that direction. Solaris extends the same underlying technology into a new category: instead of a model that generates a video of something happening, it's a model that generates and drives an interactive application in place of a traditional coded stack. Runway positions this as a challenge to the standard app-development pipeline, where design tools produce near-finished-looking mockups that still have to be manually translated into code before anything can actually run or respond to a user.

The company is not shipping Solaris as a public product yet. It's described as an early research step, gated behind a request-access form for select partners, with Runway citing open technical problems around session-length coherence, visual quality at higher resolutions, and integrating generated interfaces with accessibility tooling like screen readers.

Why it matters

If Interface World Models mature into a real product category, they represent a more radical claim than most "AI coding" tools on the market: not that AI can write the code faster, but that code as an intermediate layer for user interfaces might not be necessary at all. That's a direct challenge to the current wave of AI coding assistants (Cursor, Replit, and others) that still produce and ship conventional source code. For now Solaris is a research preview with real open problems — Runway's own write-up flags accessibility and long-session coherence as unsolved — but the user-study results against a frontier LLM's coded output give Runway a concrete data point to point to as it looks for design and enterprise partners ahead of a wider release.