Runway has introduced Solaris — the first model in the Interface World Models family, or "interface world models": within it, an app is not programmed but generated frame by frame in real time at 720p, reacting directly to clicks and drags without code or pre-described screens. The model cannot be tried yet — the company has only opened applications for early access.
What happened
Solaris renders an app interface frame by frame, like video, at 720p resolution: clicks and drags are fed into the model as conditioning for the next frame, on par with text or an image. The technical foundation is the Gen-4.5 video model: frames are generated autoregressively, multi-step denoising is distilled to a few steps, and the fast model is further trained on its own outputs to keep long sessions stable. A separate LLM handles the "intelligence": it decides whether to modify the current scene or transition to a new one and writes prompts, while the world model handles rendering. The company cites a target latency of less than 0.5 seconds — the threshold at which interactivity begins to feel natural. In a user study involving 250 people, approximately 7,500 pairwise comparisons were collected against interfaces coded by Claude Opus 5 based on the same prompts: Solaris was preferred in 61% of cases versus 24%, and in another series of comparisons, the shares were 71% versus 21%. Along with the announcement, a "translation cost" benchmark was published: 30 interfaces were reconstructed from screenshots by models up to Claude Fable 5, with visual similarity evaluated using SSIM and DINOv3 metrics. This is currently a research preview without public access.
Context
Solaris continues the path Runway began with the general world model GWM-1: there, the world model reproduces not an image but the behavior of an environment. In such an architecture, the interface lives in pixels, without a DOM or pre-described screens, and the app's behavior is determined by the world model rather than scripted workflows. This underpins the company's main argument: according to Runway's conclusion, the "translation cost" benchmark shows that translating an interface into an intermediate representation — code — irreversibly loses information, so the translation layer is proposed to be removed entirely. Notably, real-time performance here is assembled from a standard set of inference engineering techniques and openly described in the announcement — this makes the demo an honest benchmark rather than a closed showcase.
Why this matters for the industry
Solaris attacks the basic development cycle of "design → code → interface": if translating an interface into code truly loses information, tools that convert mockups and screenshots into code are put at risk, and the "code" layer and partially the "design system" layer begin to shrink — first for prototyping and one-off interfaces. A second consequence concerns computer-use agents: generated interfaces provide them with constantly changing environments for training, which did not exist before, instead of overfitting to fixed markup; synthetic worlds for such agents may emerge as a separate industry. Finally, the real-time engineering recipe from the announcement can already be factored into plans for your own products, and the "translation cost" protocol is compact and cheap to reproduce by third-party groups.
Why this matters for users
Solaris cannot be tried today by either readers or companies: access is only via an early access application through the form on the announcement page, with pricing, limits, and API surface undisclosed. The practical value right now lies elsewhere: the announcement is worth reading in full — it includes a demo and an open list of limitations. If early access expands, the first external applications will be niches where rendering errors are not critical: prototypes, demo stands, branded scenes, interactive concepts from prompts. Reasonable steps today: record product hypotheses around open scene behaviors, review the prototyping backlog, and check your own product for "translation layers" that can be removed — or that will be removed for you.
What is still unknown / limitations
The evidence base is currently typical of a vendor announcement. In the user study, there is one baseline — interfaces coded by Claude Opus 5 — and the protocol is not published; the shares of 61% versus 24% and 71% versus 21% do not sum to 100%, and the fate of the remaining comparisons, such as ties, is not disclosed. The "translation cost" benchmark is a vendor measurement of visual similarity without a published methodology: it does not prove the thesis of irreversible information loss in code, and it is correct to speak only of a direction. Actual latency, text stability, and behavior in long sessions have not been confirmed by independent measurements. The open problems that Runway itself lists are fundamental: readable text and grounding are known weaknesses of video diffusion, and compatibility with screen readers is hindered by the lack of a semantic layer in a pixel-based interface — this is a consequence of the "interface as a frame" architecture itself, not a bug in the current version. The scenario in which code as an interface layer disappears is a conditional branch on the horizon of about two years, not a completed fact.
Sources
Author
Look at AI, editorial team