Runway Solaris: The AI ​​that generates frame-by-frame interfaces (and what it still can't)

On August 31, 2026, Runway published Introducing Solaris : the first model in a family it calls Interface World Models . The question posed in the announcement is straightforward: what happens when an operating system generates apps and websites as it uses them?

Instead of translating a design into HTML, CSS, or a UI engine and then running it, Solaris synthesizes the interface frame in real time. Every click, drag, or gesture influences the next frame. There are no static screens waiting to be deployed: the visible image is the application itself.

That sounds like completely skipping front-end development. Runway doesn't market it that way. The post itself specifies early access, 720p quality, and strict limits on text, trust, long sessions, and accessibility. This article focuses on what the company documented and what that means for a team currently building digital products in Mexico or Latin America.

What is an Interface World Model (according to Runway)

Runway juxtaposes two worlds that until now lived separately:

• Systems that "know" things (search engines, assistants): respond with text, an image or an embedded video, but do not behave like a live app.
• Systems that respond in real time (JavaScript/CSS, game engines, interactive world models): are rich and interactive, but do not know their product or the user's task.

A World Interface Model, says Runway, has to be both things at once: understand the user's intent and continuously render an interactive world. Solaris treats user input (clicks, drags, other interactions) as conditioning the next frame, just as it treats text or images. A language model interprets the request, decides whether to modify the scene or switch to another, and Solaris renders how that change looks and responds accordingly.

The system relies on Gen-4.5 (its video model) and the open line with GWM-1. To achieve interactive speeds, Runway describes three steps: autoregressive frame generation, a few-step denoising distillation, and fast model training on its own outputs to stabilize long sessions. The reported visual quality remains at 720p .

What new capabilities does it announce (and which ones does it not)

Runway summarizes three properties:

• Entirely visual. If the image is the application, no secondary implementation is needed underneath. The example in the ad: a virtual showroom where you drag a garment onto yourself from a reference photo.
• Alive. The scene continues to evolve (reflections, objects that react) even if you don't press anything; you can ask in natural language to move a table or change the color of a sofa.
• Open. It is not limited to the flows that a developer anticipated in code; the same scenario can respond to different behaviors depending on the interaction.

It also positions Solaris as a training environment for agents: current LLMs, trained on hardcoded interfaces, fail when the layout changes even slightly. By collapsing visual action and response, the model can be trained against interfaces that mutate or that never existed.

What the announcement doesn't include: public release date, API, pricing, hardware requirements, or open weights. Runway is soliciting early access through a form on the same page and says it's working with partners for a public release. Any "available now for your SMB" figures you see outside of that page are not from the official post.

How Runway measured it (and what weight to give it)

In the evaluation section, Runway compares Solaris to a coded interface generated by a cutting-edge language model (Claude Opus 5 in their writing). Both started from the same image and received the same interaction requests. They then conducted a study with 250 participants , 30 examples , and nearly 7,500 peer reviews .

• Preference for Solaris when following the instruction: 61% versus 24% of the coded result ( 13% equivalent).
• Preference for "more natural" behavior on stage: 71% versus 21% ( 6% equivalent).

These figures are published by Runway based on their own study. They are not an independent production benchmark, nor do they demonstrate that Solaris can replace a checkout, a banking dashboard, or a CMS. They serve to illustrate the product's central thesis: that a generated interface can feel more coherent with the scene than a UI patch written from a screenshot.

Separately, they measure how much information multimodal LLMs lose when reconstructing interfaces from a capture (SSIM and regional similarity with DINOv3 across 30 interfaces). The message: the translation from design to intermediate representation to screen degrades fidelity; Solaris attempts to operate directly on the visual.

What it still can't (and why it matters to a business)

The post itself lists current limits. It's advisable to read them with the same seriousness as the demos.

• Text. Stable and legible text remains one of the biggest challenges in video generation; interfaces depend on it. Runway proposes hybrid approaches (pausing and rendering text-dense views with image models) as a practical solution. Fully usable, real-time generated text remains an open goal.
• Trust. In business or educational settings, a convincing but incorrect answer is worse than no answer at all. Today, Solaris anchors itself to the initial frame and reference material; conditioning with verified data throughout the session is active research.
• Long sessions. Maintaining visual and semantic consistency in open and prolonged interactions is still not resolved.
• Accessibility and integration. A generated interface must be compatible with screen readers and accessibility APIs. Visual flexibility without that foundation is not a product ready for real customers.

For a startup or SME, the useful takeaway isn't "we no longer need front-end development." It's this: live interface demos will accelerate prototyping and brand experiences, but checkout, regulated forms, SEO, performance, and compliance still rely on deterministic, measurable, and auditable software. Anyone who confuses a research preview with a production stack will waste time on the wrong demo.

What can a team do this week (without early access)

Solaris is not on their list of priorities. Even so, the announcement is pushing for concrete decisions about how to prototype in 2026.

• Separate the demo from the product. Use UI generation (v0, Figma Make, captures + LLM) to explore feel and flows. Freeze requirements, error states, gaps, and accessibility on a real front end before speaking with paying clients.
• Measure what you can control. LCP, INP, and CLS, responsive forms, legible text, and contrast: that's what sells today. A "live" interface that can't be read or audited won't close a B2B sale.
• Prepare anchor data. If Runway enters early access or faces a similar competitor in the future, they emphasize grounding the product with real product images and verified materials. Organize your catalog, typography, and branding elements now.
• Train the team in intent, not just components. If the interaction is described in natural language, the bottleneck becomes writing the task well and validating the result—the same muscle you already need with code agents.

If you're building an MVP or a store that needs to load, respond, and not invent prices, Presticorp works on that approach in startups , SMEs , and e-commerce : first a usable product, then generative interface experiments.

The writer's suggestion

Solaris is a clear sign of where the industry wants to go: less "design → code → screen" and more "intention → interactive frame." The announcement date is August 31, 2026; the status is research/early access; the reported resolution is 720p; the preference study (61%/71%) is from Runway, not a market standard.

This week, do just one thing: take the most critical flow of your product (sign up, quote, or pay) and note which parts require stable text, auditing, and accessibility. Don't delegate those parts to a world model yet. Demos can live in a design sandbox. When Runway (or another platform) opens real access with SLAs, you'll already know which surfaces you can experiment with and which should remain code.

Sources

• Runway. «Introducing Solaris». August 31, 2026. https://runway.com/news/research/introducing-solaris
• RuntimeWire. "Runway says Solaris generates interfaces frame by frame in real time." https://runtimewire.com/article/runway-solaris-interface-world-model-real-time
• TechTimes. "Runway Solaris Launches: App Interfaces Now Run as Live Video, Zero Code." September 1, 2026. https://www.techtimes.com/articles/326134/20260901/runway-solaris-launches-app-interfaces-now-run-live-video-zero-code.htm

Editorial note: Dates, Interface World Model definition, 720p resolution, study figures (250 participants, ~7,500 trials, 61%/71%), and the list of limitations (text, trust, long sessions, accessibility) are taken from the official Runway announcement consulted on September 1, 2026. Prices, GA dates, and enterprise adoption percentages are not fabricated.