Runway Unveils Solaris for Real-Time Generated Interfaces

Runway on 31 August 2026 introduced Solaris, which it calls the first Interface World Model: a real-time system that generates interactive user interfaces frame by frame instead of turning a design into code. A user study with 250 participants, 30 examples, and nearly 7,500 pairwise judgments preferred Solaris to Claude Opus 5 coded interfaces on 61 percent of instruction-following comparisons and 71 percent of natural-behavior comparisons. The research post says the model targets 720p at interactive speed, with early access by form, partner work underway, and no public launch date or price.

What an interface world model generates instead

Conventional software still splits look from behavior. Designers produce screens, then engineers recode those screens as pages, components, or engine objects so something can actually click. Runway’s 31 August research note argues that translation is a lossy step: the live product becomes a simplified stand-in for the original visual idea, and every extra behavior has to be programmed before a user arrives. Solaris is built to skip that conversion. It renders the interface itself, one frame after another, so the pixels on screen are the running application rather than a screenshot waiting for an implementation.

The model is adapted from Runway’s Gen-4.5 video system and follows the path opened by GWM-1, the company’s earlier general world model. User input — clicks, drags, typed or spoken requests — conditions the next frame the way a text prompt conditions a video model. A language model sits beside the world model to decide whether an action should tweak the current scene or jump to a new one. Runway says it converted a slow diffusion teacher into an autoregressive, distilled, self-trained engine so generation can keep up with a person instead of rendering a clip after the fact. That is a research product claim, not a shipping consumer app with a checkout button.

The pitch is also a training story for agents. Text-only models that memorize one hotel website often fail on a slightly different layout. If the interface is generated continuously, agents can practice against layouts that never existed in a training crawl. Related PromptCrates coverage of OpenAI retiring the official DALL-E GPT shows how fast generative surfaces get renamed or replaced; Solaris is trying to make the surface itself generative rather than a branded GPT wrapper around still images.

How Solaris beat coded UIs in testing

Runway first measured how much information multimodal language models lose when they try to rebuild a website from a screenshot, using structural similarity and DINOv3 features across 30 interfaces. Those reconstruction tests, including work with Claude Fable 5, are the setup for the head-to-head that matters this week: Solaris versus a coded interface produced by Claude Opus 5, starting from the same image and the same interaction request. The question was not which stack can ship a design system. It was which result better followed the instruction and which behaved more naturally inside the scene.

The 250-person study is the number to remember. Participants preferred Solaris for following the requested interaction in 61 percent of pairwise judgments, against 24 percent for the coded result and 13 percent rated equivalent. For whether the scene behaved naturally, Solaris won 71 percent, the coded site 21 percent, and 6 percent were ties. Runway’s reading is that a coded UI can often apply the requested change as an isolated update, while a world model already has a prior on how objects, lighting, and materials should move together. Those percentages are lab preferences, not a proof that generated interfaces are ready to replace checkout flows.

They do explain why a video lab is talking about operating systems instead of only clips. PromptCrates also covered Anthropic’s Claude Code weekly limit change effective 14 September; coding agents remain the default way teams ship UI, which is exactly the baseline Solaris is trying to beat on natural behavior. A creative studio that already pays for Claude Code seats should treat the 31 August paper as a competing theory of interface construction, not as a reason to cancel those seats this week.

What Solaris still cannot do reliably

Runway is unusually direct about gaps. Stable, legible text is still hard for video-like generators, yet almost every real interface depends on type. Long sessions compound small identity and layout errors. Trust is a product risk: a convincing wrong answer in a shopping or tutorial scene is worse than a blank frame. Accessibility is unsolved if the pixels are the app, because screen readers and accessibility APIs expect a document tree, not a movie. The company lists hybrid fallbacks — pause and render a still for text-heavy views — as a research path, not as a finished feature.

There is also no public price, no general availability date, and only an early-access form plus unnamed partners. Secondary coverage has mentioned a Series E of 315 million dollars; this article does not treat that figure as confirmed from Runway’s Solaris post. Buyers should read the 31 August note as a research unveil of an Interface World Model, not as a replacement for Figma, React, or a design system this quarter. The honest product status is early access by application, partner pilots, and a 720p real-time target with known failure modes.

Why generated interfaces matter for studios

If the model holds, the unit of software is no longer a pre-built app icon. A storefront could be a generated room that keeps a brand’s materials while rearranging around one shopper. A tutorial could render the next step in the learner’s actual scene instead of a fixed screencast. Creative teams already generate stills and shots; Solaris tries to generate the clickable layer those stills implied. That is why the story belongs in creative-media rather than a programming how-to: the claim is about what an interface can be, not about a new package to install.

Studios should watch three operational questions: can text stay readable at 720p real time, can a session remain coherent for longer than a demo, and can assistive tech hook into a frame stream. Until those are answered in public builds, Solaris is a 31 August 2026 research announcement with a 250-person, nearly 7,500-judgment preference study — impressive, and unfinished. File it as creative-media news about an Interface World Model, with Runway’s own research post as the primary source.

Sources

• Introducing Solaris — Runway research, 31 August 2026