On August 31, 2026, Runway published Solaris on its official research page, calling it the first model in the new generation "Interface World Model" series. It attempts to bypass traditional software transitions from design drafts to code implementation, directly generating the visual interface of applications or websites frame by frame, allowing the visuals to change continuously with clicks, drag-and-drops, and natural language commands.
This is not animating a UI image. Solaris's goal is to make the entire screen both display and interact: when a user drags furniture, the scene continuously rearranges; when a material is changed, subsequent frames retain object relationships and continue to respond. Runway believes this dynamic environment can also be used to train computers to operate agents, so they no longer just remember the button positions on a fixed webpage.
Reasoning and rendering are split into two layers
Solaris's Gen-4.5 video generation model adaptation based on Runway also continues the GWM-1 world model approach. In the system, the language model understands user intent, determines whether a scene is changing locally or as a whole, and provides the next action; The world model renders decisions in real time into continuous images. Actions like clicking and dragging act like text and image prompts, becoming conditions for the next frame.
To reduce video diffusion from "seconds to minutes to generate a segment" to an interactive speed, Runway uses autoregressive frame-by-frame generation, distilling multiple denoising steps into a few steps, then training fast models on their own outputs to maintain long-term visual stability. Current official goals include real-time response, in-session consistency, and 720p visual quality. To understand the difference between this approach and ordinary video generation, check out the article on "Why World Models Are the Key Capability for AI Agents" (/article/1301-world-model-why-its-considered-the-next-key-capability-for-ai-agents-and-bots).
The official review highlights the advantages but also leaves questions marked
Runway invited 250 participants with 30 interaction samples and collected nearly 7,500 paired judgments. Compared to the code interface generated by Claude Opus 5, Solaris received a 61% preference for "whether to follow interaction instructions," while the code scheme was 24%; For "whether behavior is natural," Solaris scored 71% and the code scheme for 21%. This is a review designed and published by Runway itself, demonstrating potential for visual continuity but not directly equating with software reliability, performance, or maintenance costs in production.
At this stage, it's more like an interactive prototype, not a code substitute
Solaris still has four clear boundaries. The hardest text to stabilize in generating videos happens to be the interface most relied upon; Incorrect but realistic teaching or commercial visuals pose trust risks; Objects, semantics, and states in long conversations may gradually drift; Generative interfaces also need to integrate screen readers, accessibility APIs, and existing software systems. The cost of frame-by-frame generation is still higher than delivering a pre-built webpage.
Therefore, Solaris is currently more suitable for immersive product displays, variable teaching demonstrations, proof of concept, and agent training environments, rather than directly carrying highly deterministic processes like payments, account setup, or precise forms. Runway is currently working with partners to promote public release, with pages only open for early access applications. What is truly worth watching is not whether "the code will disappear," but whether visual generation can cross the minimum barriers of software engineering in terms of text accuracy, state traceability, and accessibility support.
