Runway has released Solaris, an Interface World Model that generates working applications by predicting pixels frame-by-frame in response to user interactions, rather than writing HTML, CSS, or JavaScript code. Announced on August 31, 2026, Solaris represents a fundamentally different approach to AI-powered app building compared to existing tools like v0, Lovable, and Bolt, which all produce code as an intermediate step.

What Makes Solaris Different From Code-Based UI Builders?

The distinction between Solaris and traditional code-generating tools is not just technical; it changes how the entire system works. Code-based generators take a prompt and produce HTML, CSS, and JavaScript files that a browser then parses, lays out, and executes to create the visual interface. Solaris removes that middle layer entirely. Instead of generating markup, the model predicts the actual pixels you see, frame by frame, directly from your interactions. There is no code being generated at any point, and nothing a browser is executing.

This architectural choice has real trade-offs. Code-based tools produce inspectable, editable source files that developers can read, modify, check into version control, and hand off to teammates. Solaris outputs live video frames with no underlying code to inspect or edit. If you want to change something, you re-prompt or re-interact with the interface; there is no line-level code editing. Runway targets 720p output with sub-500-millisecond per-frame latency, fast enough to feel responsive rather than laggy.

How Does Solaris Actually Work Under the Hood?

Solaris is built on Runway's existing Gen-4.5 video generation model, adapted in three specific ways to handle interactive app generation. First, it uses autoregressive generation, meaning each frame depends only on the frames before it, which enables real-time, open-ended interaction instead of generating a fixed-length video clip. Second, it uses compressed denoising, collapsing the normally many denoising steps per frame down to just a few steps to hit interactive latency targets. Third, it splits reasoning between a language model, which handles interpretation and behavioral logic (what should happen when you click "submit"), and the world model itself, which handles the actual visual rendering of the next frame.

A world model, in technical terms, is a neural network that learns the dynamics of an environment by predicting what happens next given the current state and an action. Instead of predicting how a thrown ball falls or how a 3D game world responds to player input, Solaris predicts how a button press should change the next frame of a dashboard, form, or application interface.

What Are the Key Advantages and Limitations?

Runway's own user study, conducted with 250 participants who made 7,500 judgments, found that Solaris was preferred over coded interfaces 61% of the time for following instructions and 71% of the time for feeling natural to interact with. When large language models (LLMs) like Claude, GPT-4o, and Gemini 2.5 Pro were asked to reconstruct an interface by describing it in language first, information loss increased with visual complexity across every model tested.

However, Runway itself flags several significant limitations that remain unsolved. These include the following challenges:

• Text Rendering: Generating legible, correctly spelled text in interfaces remains difficult and unreliable.
• Trust and Hallucination: The model can produce visually convincing but functionally incorrect output, creating a "convincing wrong" problem where the interface looks right but does not work as intended.
• Long-Session Coherence: Maintaining consistent interface state and behavior over extended user sessions is not yet solved.
• Accessibility Integration: Screen readers, keyboard navigation, and other accessibility APIs need separate integration work that has not yet been addressed.

Is Solaris Really the First Interface World Model?

Runway brands Solaris as the first "Interface World Model" aimed at general operating-system-style apps, but the underlying technique of using a world model to generate interactive, screen-like visual state frame-by-frame in response to user input has prior art. Google DeepMind's Genie 2 generates playable, responsive environments from a single image, predicting the next visual state from an action using the same core loop Solaris uses, just applied to game worlds instead of app screens. Odyssey's Starchild-1 generates synchronized audio-video in real time while responding to continuous input, making it architecturally the closest sibling to what Solaris does. Runway's own GWM-1, which powers Runway Characters, already demonstrated real-time autoregressive frame generation responding to live input months before Solaris.

So the honest assessment is that Solaris is a new application of an established world-model technique to a new domain, not a fundamentally new kind of model. That is still a meaningful step, because nobody had pointed this specific machinery at "operating system that generates apps as you use them" before, but the "first" claim is more about category naming than technical novelty.

How to Evaluate Solaris Against Existing Tools

• Output Format: Code-based tools like v0, Lovable, and Bolt produce HTML, CSS, and JavaScript files that are inspectable and editable; Solaris produces pixels directly with no underlying code.
• Version Control and Collaboration: Code-based tools integrate with standard git workflows and can be handed off to other developers; Solaris requires re-prompting or re-interaction to make changes, with no traditional version control.
• Runtime Requirements: Code-based tools rely on a browser to parse and execute the code; Solaris requires the model running live per session, which has deployment and cost implications.
• Accessibility Support: Code-based tools inherit accessibility features from real DOM elements; Solaris needs separate accessibility integration work that Runway has flagged as unsolved.

What Does This Mean for the Future of App Development?

Runway's pitch is that the intermediate step of code generation is exactly where fidelity gets lost. By predicting pixels directly, the model can render any visual concept as an interactive interface without first translating it into a UI framework's components. However, this approach trades away the inspectability, editability, and deployability that developers currently rely on. Solaris is currently available only through early access via a request form, with no pricing disclosed and no independent verification of Runway's benchmark claims.

The broader implication is that world models, which have proven effective for video generation, game environments, and avatar animation, are now being applied to a new frontier: the interactive software interfaces that billions of people use every day. Whether this approach will replace code-based generation or coexist alongside it remains an open question, but Solaris demonstrates that the underlying technology is mature enough to attempt it.