Build · founder · 7 min read

Runway's Solaris Skips the Code Entirely. What That Actually Means.

Runway says an interface can be generated frame by frame with no code underneath. Here's what's real, what isn't, and what to do about it today.

Published September 1, 2026 ·

runway no-code world models design research

Every tool covered on this site does fundamentally the same thing. You describe what you want, and the tool writes code. Lovable writes React. Bolt writes React. Replit writes whatever it likes. The code is the product; the description is just the input.

On August 31, Runway announced something that removes the code.

Solaris is what Runway calls an Interface World Model. There is no HTML, no component tree, no JavaScript running underneath. The interface is generated as video, one frame at a time, and your clicks and drags become inputs that shape the next frame. Runway’s framing is deliberately large: what happens when the operating system generates apps as you use them?

Before you rearrange any plans around this — you cannot buy it, there is no price, and there is no launch date. More on that below. But the idea is worth ten minutes of your attention, because it attacks an assumption underneath every product decision you make.

The assumption it attacks

Runway’s argument goes like this. A design has to be translated into an intermediate representation — code — before it can do anything. That translation is lossy in two directions.

It loses visual fidelity, because a rich design gets reduced to boxes and rules that a browser can render fast.

And it loses possibility. Every behaviour has to be defined ahead of time by someone who guessed what users might do. Runway’s phrase for shipped software is “a lossy compression of the space of possible interactions, frozen before any user arrives.” That is a sharper description of the product-spec problem than most product people manage.

Solaris generates the pixels and the behaviour together. Nothing is defined in advance, so nothing has to be anticipated.

How it works

Two models, split by job.

A language model reads what you did — a click, a drag, a typed request — and decides what should happen next. A world model, built on Runway’s Gen-4.5 video model, draws the result. The LLM handles reasoning; the world model handles rendering.

The engineering problem was speed. Interactions stop feeling interactive at roughly half a second of delay, and video diffusion models take seconds to minutes. Runway got there by generating frames one at a time rather than refining a whole clip, distilling the denoising process down to a few steps, and retraining the fast model on its own output so quality holds over a long session. It targets 720p.

One consequence is genuinely novel. Because interactions are described in language rather than programmed, any object can become a tool. Click a cat, and your next clicks paint its fur onto whatever you touch. Nobody built that feature.

The benchmark worth reading

Runway ran two evaluations. The second one got the headlines; the first one is more interesting.

The translation test. Runway took 30 website interfaces and asked state-of-the-art multimodal models — GPT-4o, Gemini 2.5 Pro, Claude Fable 5 — to recreate each one from a single screenshot, then measured how much survived. Every model lost information, and the loss got worse as visual complexity rose. Natural images fared worst, because rich visual detail does not survive a trip through language.

If you have ever handed a Figma file to a design-to-code tool and gotten back something that was technically the same layout and visibly not the same thing, this is that, measured.

The head-to-head. Runway put Solaris against a coded interface generated by Claude Opus 5, same starting image, same interaction requests, and ran a study with 250 participants across 30 examples — nearly 7,500 pairwise judgments. Solaris was preferred 61% to 24% on following the instruction, and 71% to 21% on behaving naturally in the scene.

Treat those numbers as what they are: a vendor’s own study, published by the vendor, with no independent replication. The larger gap on “natural behaviour” is the honest signal, and it matches the argument — a coded interface treats an interaction as an isolated state update, while a world model already knows how objects and materials behave.

What it can’t do, in Runway’s own words

The launch post has a section titled “What It Can’t Do Yet,” which is more candor than most launches offer. Four limits, and two of them are disqualifying for a real product today.

Text is unstable. Legible, stable text is one of the hardest problems in video generation, and interfaces are made of text. Runway floats a hybrid where image models render text-heavy views and video handles motion. That is a workaround, not a solution.

A convincing wrong answer is worse than no answer. Runway’s words. A generated interface can render a product detail, a price, or an instruction that is plausible and false. Right now the scene is anchored by the starting frame you supply. Grounding it in verified product data is described as “an active research focus,” which means it is not solved.

Long sessions drift. Coherence over extended use is still open research.

Accessibility is unsolved. A generated interface still has to work with screen readers and accessibility APIs. Nothing exists here yet. For anyone shipping to real customers — particularly in regulated sectors, or anywhere subject to accessibility law — that alone rules it out of production for now.

What you can actually do today

Nothing. Solaris is not a product. There is no pricing, no public availability, no launch date. Runway says it is “working with key partners” and is collecting requests through an early-access form.

So the practical answer is: fill in the form if you’re curious, and change no plans.

The reason to pay attention is the direction, not the release. Three things follow if this line of work matures:

• The app stops being the unit. If the system can generate whatever interface your task needs, there’s less reason to package software into a fixed catalogue of apps you install and open.
• A storefront stops being one layout. It becomes a generated environment that keeps your brand identity while reshaping around each visitor.
• The design-to-code handoff stops existing. Not automated — removed.

The honest read

Solaris is a research demo with an unusually good argument attached and unusually honest limitations. It is not a threat to your roadmap this quarter or next.

It is a reason to be sceptical of anyone selling you a five-year platform bet. The thing every vibe coding tool currently sells — the translation from your intent into code — is being framed by a serious lab as the step to delete. That may not work. But when you’re choosing what to build on, prefer the tools that let you leave. See picking AI tools that last and what is vibe designing for the version of this question you can act on this month.

---

Sources read directly: Runway — Introducing Solaris (August 31, 2026). Corroborating coverage: The New Stack, The Rundown AI.