共有:
Interface World Model

The video-AI company Runway
now generates the software itself.

On August 31, 2026, Runway unveiled Solaris, a new system built on its Gen-4.5 video-generation model. Instead of writing screens as code, it renders the response to every click and drag frame by frame, in real time — a new kind of model that generates a software interface the way it would generate a video.

AI Navigate Editorial2026.09.027 min read
CODE GENERATION Prompt LLM writes code Build ↓ screen Every change needs a rebuild SOLARIS Click Next frame, generated directly No compile step
01
The Bottleneck

Why software has always
had to pass through code

Until now, getting AI to build a software screen meant, in practice, getting AI to write code. Coding assistants like Claude and GitHub Copilot turn a prompt into source code, which then has to be built and compiled before it becomes anything you can look at. That detour through code as an intermediate representation is exactly what forces a rebuild after every single change.

Runway's newly announced Solaris removes that step. According to Runway's own announcement, Solaris is the first entry in a new model category it calls an "Interface World Model." Built on the Gen-4.5 video model, it takes a user's clicks and drags as input and generates the resulting screen frame by frame, directly. There is no source-code artifact in between — the generated video is the software.

Code-generation AISolaris (Interface World Model)
Output is source codeOutput is a screen frame (video)
Requires a build/compile stepRenders the generated frame directly
Every change means rewriting codeGenerates the next frame from user input
Example: Claude Opus 5 writing codeRunway Solaris generating frames

An app no longer runs as code — it runs as video.


02
How It Works

It watches your input, then predicts the next screen

Runway carried the video-generation know-how from Gen-4.5 into the lineage of its game-like world model, GWM-1, and pointed it at the screen itself.

01

Feed in the user's action

A click, a drag, a scroll — each becomes a cue for what the model should draw next. There's no pre-built set of screens to navigate between, unlike a conventional UI.

02

Generate the next frame directly

A single model built on Gen-4.5 renders, frame by frame, the screen that should follow from that action. Skipping the code intermediary means there's no build step to wait on.

03

The generated video becomes the UI

The result isn't a preview image — it's an interactive screen you can act on. When the user acts again, the model predicts and generates the next frame in turn.

03
The Numbers

How it compared against code-generation AI

In Runway's own comparison study, testers preferred Solaris's output over screens coded by a coding assistant more often than not.

71%
Preferred over Claude Opus 5-coded pages on in-scene behavior
61%
Preferred on instruction-following accuracy
Early access
Currently a research preview, not a public release

Per Runway's own report, testers picked interfaces generated by Solaris over pages coded by Claude Opus 5 in 71% of matchups on in-scene behavior, and in 61% of matchups on instruction-following. This was an internal Runway study, though — it hasn't yet been reproduced by an independent third party, which is worth keeping in mind.

04
Who It Hits

Whose workflow actually changes

Engineers

You can try out UI ideas interactively without writing code. But this is a research preview today — not something to wire into a production product yet.

Designers

You may be able to try many interactive prototypes — including how they feel to use — before anything gets implemented, pushing the "touchable" stage earlier than wireframes allow.

Product managers

Concept validation could get faster, but with no launch date or pricing announced yet, it's too early to put this on any roadmap.


05
Frontier

From "a video company"
to "a computing company"

Runway has consistently been described as a video-generation company. What makes Solaris significant is that it redirects the company's video-generation know-how (Gen-4.5) and its lineage of game-like "world models" (GWM-1) into an entirely different domain: software UI. The ability a video model develops to "predict and render what happens next" turns out to generalize beyond video, to computer screens in general.

By role, designers and prototyping-focused engineers stand to benefit first — the time it takes to turn an idea into something you can actually interact with could shrink substantially. For product managers, faster concept validation is a real upside, but it's still too early to build it into any decision. The sensible action right now is to follow Runway's progress through its early-access request form, while leaving existing prototyping workflows untouched for the time being.

There's a real counter-case, too. First, the published numbers (71%/61%) come from Runway's own evaluation, not an independent third-party replication. Second, this is currently an early-access research preview — no public launch date or pricing has been announced. Third, it remains to be seen how far generated screens can hold up against a real production system — database and API integration, accessibility, and so on — once tested beyond a demo. "No code required" is a compelling pitch, but it's still a long way from replacing how software actually gets built.