Runway’s Solaris tests whether live AI video can become an interface

Runway introduced Solaris on August 31 as an early preview of websites and apps rendered as live, interactive video. At the time of The Rundown’s September 1 coverage, the company was inviting requests for early access. Its central idea: generate what an interface does and shows as someone interacts with it, without a separately coded implementation of that interface.

Runway calls Solaris an “Interface World Model.” Its announcement describes a system that draws successive frames in response to clicks and drags. That opens an intriguing possibility for the web, though the preview leaves open whether these experiences can become useful, dependable, and affordable.

The launch was originally covered in The Rundown’s September 1 newsletter.

How Solaris turns interaction into video

Solaris pairs Runway’s Gen-4.5 video model with a language model. The language model interprets a user’s action, decides what should happen, and prompts the video model to render the result. Runway’s examples include dragging a shirt from a virtual rack onto a photograph, assembling a salad by dropping in ingredients, and interacting with a combustion demonstration.

These examples make the appeal concrete: an object in a scene can respond directly to someone’s actions. For a shopper exploring clothing or a learner examining a process, that could create a more immediate visual experience.

A technical paper submitted September 1 adds detail about the engineering. It describes generating frames sequentially, reducing the number of generation steps through distillation, and training the model on its own outputs. Those methods help explain how Runway is pursuing speed while trying to preserve quality across successive interactions.

The limitations are substantial. Runway acknowledges difficult text rendering, drift during longer sessions, and screens that look convincing while showing something wrong. The paper also identifies unresolved questions around accessibility and integration. A successful visual demonstration therefore leaves considerable work before a service could depend on the interface.

Why it matters

Solaris could be a glimpse of a new kind of web experience, or it could remain a novelty. The larger significance lies in the possibility that improvements in model speed, cost, and quality together could make previously impractical interfaces feasible. That argument depends on continued progress across all three.

Speed matters because it changes what people can do with generated imagery. When frames arrive during an interaction, someone can steer an experience as it unfolds. There are signs of this direction beyond Solaris: Google DeepMind reported navigable generated worlds at 720p and 24 frames per second with Genie 3 in August 2025, while acknowledging limits on text, actions, and session duration.

Smooth animation alone cannot establish responsiveness. In its May 2026 Characters engineering account, Runway reported 24 frames per second alongside 1.75 seconds of server response latency after someone stopped speaking. Those measurements belong to a different product, but they illustrate why Solaris needs evaluation across the full interval between an action and its visible result.

For designers, retailers, and educators, short visual experiences seem a plausible starting point. Exploring a garment or manipulating an instructional scene could benefit from direct visual feedback. Whether those benefits survive repeated interactions needs testing. Text legibility and session drift make longer workflows harder to trust, while a plausible but incorrect instructional scene could mislead a learner.

The economic case is equally conditional. Runway says Solaris costs substantially less to run than standard video diffusion, while acknowledging that continuously generating frames still costs more than serving an already built page. A custom interactive experience could justify that recurring expense if it delivers enough value. Published per-session costs and evidence of that value are still needed to judge the tradeoff.

Quality also includes dependable behavior, readable information, accessibility, and connections to other software. A shirt appearing on a photograph demonstrates a visual interaction; a shopping service also needs accurate product information and reliable integration. Solaris’s unresolved grounding and integration questions make that distinction consequential for anyone considering deployment.

The same model improvements could also strengthen conventional software. Anthropic’s July announcement of Opus 5 claimed stronger capabilities at unchanged token prices, including improvements in coding. Generated interfaces will be evaluated alongside increasingly capable coded experiences and potential combinations of the two.

Solaris puts the writer’s central possibility into view: if speed, cost, and quality keep improving together, builders could attempt interfaces that previously required too much time or money. Its practical prospects depend on how quickly it responds, how reliably it behaves, and what each useful session costs.