TL;DR
Solaris renders an interface continuously from a world model instead of assembling it from components, which lets the interface adapt to whoever is using it. That adaptivity is the interesting part, especially for accessibility. Runway lists accessibility among the things it has not solved. It is early access behind a request form, so nobody outside Runway’s partners can test any of it, and it could be a flash in the pan.
In plain English
A world model is trained to predict what a scene looks like next, given what it looks like now and what just happened to it — the same family of model behind generated video. An intermediate representation is the code a design gets turned into so a browser can run it: HTML, CSS, JavaScript. Runway’s claim is that Solaris removes that step. The screen is generated directly, and your click is an input to the next frame rather than an event handled by code.
A user interface can be rendered the way a video model renders video: frame by frame, in real time, responding to clicks and drags as they happen. No DOM, no component tree, no code underneath. That is Solaris, published by Runway on 31 August as “the first model in a new family of AI systems we call Interface World Models,” with access by request form. Their argument is that turning a design into code before it can do anything throws information away, and that skipping the step avoids the loss.
Runway’s announcement on X described it this way:
“Solaris is a new kind of operating system that generates interactive interfaces frame by frame, in real time, with no code. We find that Solaris outperforms frontier LLMs when generating new interfaces, across structural similarity and information retention.”
The research page reports two separate experiments.
The first is a reconstruction benchmark. Runway took 30 interfaces and asked multimodal language models, Claude Fable 5 among them, to recreate each one from a single screenshot, scoring the results on structural similarity and on whether visual content survived using DINOv3 features. Every model tested lost information, and Runway reports that reconstruction quality “consistently degrades as visual complexity increases.” That benchmark measures models copying an interface that already exists.
The second is a preference study. Runway compared Solaris against Claude Opus 5, starting both from the same image with the same interaction requests, and asked 250 participants for roughly 7,500 pairwise judgements. Solaris was preferred in 61% of comparisons on following the instruction against 24% for the coded result, and 71% against 21% on behaving naturally within the scene. That is a wide margin on the second measure.
Solaris is aimed at something people already try to do: make software that adapts to the person using it. Runway’s examples are storefronts that reshape around intent while keeping the brand recognisable, and tutorials that render the next step in your context rather than replaying the same sequence for everyone. Nobody can say yet whether it works, but it is clear what it is meant to do.
What piques my interest is what a generated interface could mean for accessibility.
An interface that reshapes itself around the person using it is the opposite of the one-size-fits-all screen that already excludes people. We should not be gatekeeping how people use these tools. The technology meeting someone where they are, rather than requiring them to meet it, is the argument I made about writing with AI as an accessibility tool, and generated interfaces are a much larger version of it.
Runway puts it on the list of things Solaris cannot do yet:
“Accessibility and integration. A generated interface still needs to work inside the rest of the software stack, including assistive technologies such as screen readers and accessibility APIs, so that flexibility doesn’t come at the expense of usability.”
That is honest, and it is a serious problem. A screen reader works by traversing a structure: elements, roles, labels, an order things sit in. Solaris removes that structure on purpose — it is what the model is designed to do. An interface with no DOM has nothing to traverse, and generating pixels that look like a button does not produce anything a screen reader can announce. Runway also lists text rendering as an open problem.
So the technology with real promise for adapting to individual needs currently does not work for people who rely on screen readers and other assistive technology. Both things are true, and I am not going to pretend one cancels the other.
If the interface is generated rather than built, what does the designer specify? Is it still the steps a user takes through a flow, or does it become the win condition — describe what winning looks like, and let the model work out the interface that gets there?
I am putting that as a question because I do not know, and neither does anyone outside the early access programme. It is what I would want answered before deciding whether this is where interface design is going.
What prevents misinformation and abuse here? Guardrails, which is what I wanted to see from Anthropic’s Model Hardware Standard last week too.
Solaris removes the artifact you would normally inspect. When an interface is code, the code can be read, diffed, reviewed and held up as evidence of what it was built to do. When the interface is generated frame by frame with no intermediate representation, there is nothing to read. All you can inspect is what appeared on screen while you were using it. How do we know its intention?
Runway acknowledges that “for instructional or commercial experiences, a convincing wrong answer is worse than no answer,” and their current answer is grounding: the starting frame is composed from real product imagery and reference material, and conditioning on richer verified context is described as an active research focus. Runway calls this active research, which means it is not solved.
Of the three barriers Runway names — speed, staying coherent across a session, and cost per frame — coherence matters most to me, then speed, then cost. Runway leads with speed and the half-second threshold where interaction stops feeling interactive.
This is an interesting foray into how things could go right and how they could go wrong, and I do not think anyone can tell you yet which way it lands.
The proof of the pudding is in the eating. It is early access behind a form, so nobody outside Runway’s partners can try it yet (but I’ve put my hat in the ring).