← Back to feed

Runway shipped Solaris on August 31, 2026 — the company’s first Interface World Model, and the first production system to generate interactive applications entirely as rendered frames rather than as code executed against a DOM.

The architectural argument is simple: every piece of software built today requires a translation step. A designer produces a pixel-perfect visual. An engineer converts that into an intermediate representation — HTML, CSS, JavaScript — that a browser can parse. That translation is lossy. Behaviors have to be anticipated and hardcoded before any user arrives. The interface ships frozen.

Solaris removes the intermediate representation. A single world model generates each frame and handles each user interaction directly. No code is produced, no parse tree is maintained, no event listener is registered. The frame is the application.

Three Capabilities Runway Claims Are New

Entirely visual. Because there is no backing code layer, the visual fidelity of the design carries directly into the interactive experience. Runway’s example: a virtual clothing store where picking up a shirt and dragging it onto a reference photo of yourself is a native interaction, not a scripted behavior.

Alive. The interface renders continuously rather than waiting for a discrete user event. Lighting shifts. Objects respond to manipulation. A user can say “Move the table so I can see how it looks” and the scene updates in response — not by triggering a pre-written animation, but by generating new frames.

Open-ended. Traditional interfaces support only the interactions a developer anticipated. Solaris, Runway argues, is bounded only by the capability of the driving world model. New interaction patterns emerge without a deployment.

The Agent Training Angle

The most technically interesting claim in the announcement is about agent training, not consumer UX.

LLMs trained for computer use tend to overfit to specific interface layouts. Move a button, rename a field, redesign a page, and agent success rates drop sharply. Research cited in the Solaris announcement found that even frontier LLMs struggle on basic computer use tasks — booking a hotel, ordering groceries — when the interface differs slightly from training.

Solaris addresses this by generating interfaces that have never existed before. An agent trained against Solaris-rendered environments learns to interact with the underlying intent of an interface, not its pixel coordinates. Whether that translates to real-world generalization is unproven, but the framing is coherent.

User Study Numbers

Runway ran a preference study comparing Solaris against Claude Opus 5 producing coded output. 250 participants evaluated 30 interaction examples, generating approximately 7,500 pairwise judgments on two questions: which result better follows the instruction, and which behaves more naturally within the scene.

Runway reports Solaris was preferred on instruction-following, and that the gap was larger on natural behavior — 7x% preferred Solaris versus 21% for the coded website, with 6% equivalent. Specific instruction-following percentages were not published.

Context

Runway raised $315 million at a $5.3 billion valuation earlier in 2026, with world model expansion into new domains explicitly the stated use of funds. Solaris is the first public artifact from that work. Prior Runway world models targeted video — physics, motion, 3D environment generation. Solaris extends the same generative approach to software interfaces.

Early access is available via form at runway.com. Partner integrations are in progress; no pricing or latency figures were disclosed. No model weights are being released.