Runway announced Solaris on 31 August 2026, the first model in a family the company calls Interface World Models. The system generates the user interface directly as an image, frame by frame, instead of translating a design into HTML, CSS and JavaScript. It removes the code layer between design and working software.
Key takeaways
- Solaris builds on the Gen-4.5 video model and extends the GWM-1 world model line
- Clicks and drags condition the next frame rather than trigger a programmed event
- Study with 250 participants and nearly 7,500 pairwise judgments: 61% preferred Solaris, 24% the Claude Opus 5 coded interface
- Runway states it holds visual quality at 720p across a session
An interface without the middle layer
Software today needs a translation — a visual design must become code before it does anything. Runway argues the conversion is lossy, locking in only behaviours anticipated before launch. In Solaris the whole frame is the interface.
How Runway reached real time
Solaris builds on the Gen-4.5 video model, following the direction set by GWM-1. Clicks and drags condition the next frame, so responses need not be coded.
Speeding that loop up to real time took three stages:
- moving to autoregressive frame generation — one frame after another
- distilling multi-step denoising?Denoising distillation: A training technique in which a fast model learns to reproduce the result of a multi-step denoising process in only a few steps. down to a few steps
- training the fast model on its own outputs.
The three stages do not change the loop itself — a user gesture enters the model as a condition, and the model answers with the next frame. They only shorten the time that loop takes to close.
A separate language model runs the decision layer, settling what happens while Solaris renders how it looks.
What the tests showed
Runway put Solaris against an interface coded by Claude Opus 5. Both started from the same image and got identical instructions, with 250 participants delivering nearly 7,500 judgments across 30 cases.
| Interface | Overall preference | Natural behaviour |
|---|---|---|
| Solaris | 61% | 71% |
| Claude Opus 5 coded interface | 24% | 21% |
The margin is wider on natural behaviour in the scene than on instruction following. A separate test showed multimodal models, Claude Fable 5 among them, rebuild an interface from a screenshot worse as image complexity rises — measured with SSIM and DINOv3.
Why it matters
Runway is not aiming at designers alone. Computer use agents train on fixed page layouts and lose accuracy the moment a layout changes. An interface generated on the fly gives a training environment where the same task looks different every time. If that holds, it shifts the centre of gravity from generating code to generating visual state — a different class of problem than Copilot or Cursor.
What next?
- Runway is collecting early access requests and plans a public launch with partners
- Stable, legible text remains unsolved — Runway is weighing a hybrid setup with an image model for text-heavy views
