Contrary to traditional digital design, the modern AI-assisted workflow involves broad, conceptual exploration on canvas-like tools (e.g., Paper) and sweating the final visual details directly in code. Pixel-nudging in design software like Figma is becoming obsolete for last-mile fit and finish.

The debate over designing in code versus a visual canvas is outdated. The modern workflow isn't about choosing one, but fluidly moving between both tools based on the task: canvas for broad exploration and code for deep, interactive prototyping.

Historically, design workflows moved from low-to-high fidelity due to tool constraints. AI tools like Codex remove these barriers, allowing designers to begin with functional wireframes in code for immediate interaction testing, bypassing static sketches.

The team developed a dedicated GUI for Codex because TUIs are limiting for multimodal interactions (voice, images). They believe the ideal interface for AI programming is a GUI, but not a traditional IDE, creating a new "command center" for agents that has a higher ceiling for future capabilities.

AI makes iterating in code as inexpensive as sketching in design tools. This allows teams to skip low-fidelity wireframes and start with functional prototypes, blowing up traditional, linear development processes and reinventing workflows daily.

At OpenAI, the development cycle is accelerated by a practice called "vibe coding." Designers and PMs build functional prototypes directly with AI tools like Codex. This visual, interactive method is often faster and more effective for communicating ideas than writing traditional product specifications.

At OpenAI, the first question is "Can we solve this with the model (tokens) instead of pixels?" This treats the AI as the primary design material, pushing designers to think about interaction and behavior before creating bespoke user interfaces.

OpenAI is developing a "dynamic user interface library" designed so the AI model can interpret and compose UI elements itself. This forward-thinking approach anticipates a future where the model assembles bespoke interfaces for users on the fly.

The new GPT Image 2 model demonstrates a significant leap in capability by generating complex, structured layouts like multi-panel brand kits. Its ability to organize distinct elements and render clean typography on a single canvas makes it a powerful tool for creating sophisticated graphic assets beyond single-subject images.