Runway Solaris: An AI That Renders Apps as Video

Runway announced Solaris on 31 August 2026, and it does something no app builder currently does: instead of writing code that a browser then renders, it generates the interface itself, frame by frame, as you interact with it. Runway calls it an Interface World Model. There is no HTML, no component tree, no repository. There is a model producing pixels in response to your clicks. That distinction matters more than the demo does, and it is worth understanding before the launch coverage settles into "AI builds apps now".

What Runway actually shipped

Solaris is built on Gen-4.5, Runway's video generation model. A video model already knows how one frame should follow the last. Solaris adds a feedback loop so that the next frame follows your input rather than a prompt's trajectory. Click a button in the generated interface and the model renders what it thinks should happen next.

Runway published a user study alongside the research post: 250 participants, 30 interaction examples, close to 7,500 pairwise judgements. Solaris was preferred in 61% of comparisons against 24% on following instructions, and 71% against 21% on natural behaviour. Those are the company's own numbers from its own study, which is normal for a research announcement and worth reading as such.

It is not publicly available. Runway is taking early access requests and says it is working with launch partners.

Why this is a different category from an AI app builder

An AI app builder writes code. You get files. You can read them, host them, hand them to a developer, and keep them if the vendor disappears. That is the whole basis of the vendor lock-in question people ask before committing to one.

Solaris produces no artefact. The interface exists only while the model is generating it. That has three practical consequences:

• Nothing persists by default. A form that looks like it saved your data did not save anything, unless something outside the model did.

• Nothing is auditable. You cannot diff a frame. You cannot review what changed between two versions of a generated screen.

• Nothing is portable. There is no export path, because there is nothing to export.

None of that makes it useless. It makes it a different tool with a different job, closer to a live prototype or a simulated environment than to a product you ship.

The limitations Runway names itself

The announcement is unusually direct about what does not work yet. Stable, legible text is described as one of the hardest remaining problems, which for an interface model is a large caveat: interfaces are mostly text. Runway also lists coherence over long open-ended sessions as an active research area, and flags accessibility and assistive technology integration as unresolved.

That last one deserves a sentence. Screen readers work by reading a document structure. A generated video frame has no structure to read. Until that is solved, anything built this way is inaccessible by construction, not by oversight.

What to do with this if you build things

Almost certainly nothing this week, because you cannot get access. The useful move is to file it correctly.

If you are prototyping and want to feel how a flow behaves before anyone writes code, this class of tool is genuinely interesting. If you are building something a customer will log into, the questions that decide the tool are persistence, auditability, and accessibility, and Solaris currently answers none of them.

The wider pattern is worth noting too. World models have been discussed mostly as robotics and simulation research. Applying one to user interfaces is a real shift in framing, and if it works it moves the frontier of what "generated software" means. If you want the underlying concept rather than the product news, we have a separate explainer on what an AI world model is. For deciding whether any given launch deserves your attention at all, the release-is-it-a-big-deal checklist applies here as well as anywhere.

FAQ

Is Runway Solaris available to use?

Not publicly. As of the 31 August 2026 announcement, Runway is accepting early access requests and says it is working with launch partners. There is no general availability date.

Does Solaris generate code?

No. That is the central difference. It generates the interface directly as rendered frames. There is no source code, component tree, or exportable project behind what you see.

Is Solaris a replacement for an AI app builder?

Not as it stands. App builders produce code you own and host. Solaris produces a live generated interface with no persistence layer and no export. They solve different problems.

What model is Solaris built on?

Runway's Gen-4.5 video generation model, adapted so that the next frame responds to user input rather than following a text prompt.

Why does text legibility matter so much here?

Interfaces are overwhelmingly made of text: labels, values, error messages. Runway names stable legible text as one of the hardest open problems, which limits how far the current approach can go for real applications.

Sources: Runway, Introducing Solaris and Runway's announcement thread, both published 31 August 2026. If you want to keep track of releases like this without reading every announcement, our guide on how to keep up with AI news covers a workable routine.