Artificial intelligence coding agents are specializing faster than anyone expected. A new test pushing three major AI developers to build complex, graphic-heavy web interfaces proves that not all code generators understand visual design. Claude Code handled the task flawlessly. The competition did not.
The results come from a highly detailed XDA Developers benchmark evaluating the frontend design capabilities of Claude Code, Codex, and Google Antigravity. The tools were tasked with building a complex project management app named TaskForge and a high-end sneaker launch website. The goal was to test UI/UX design judgment rather than basic background code generation.
Claude Code emerged as the clear winner. The agent successfully handled small design details, spacing, and fluid typography. It felt like a seasoned human developer making active aesthetic choices. Google Antigravity 2.0, powered by Gemini 3.5, produced a clean, balanced, and fast prototype. But it missed nuanced UX touches.
OpenAI Codex, currently running GPT-5.5, was the weakest in visual execution. It delivered a functional but entirely unpolished wireframe riddled with scaling issues, small fonts, and naked layouts.
Why UI Nuance Separates the Agents
The gap in visual frontend execution highlights the distinct personas these models have developed by mid-2026. Claude Sonnet 4.6, released in February, was heavily tuned for scoped, repo-level execution. It completes end-to-end tasks exactly as asked.
This aligns perfectly with the broader engineering consensus on model release dates and the specific strengths of each agent regarding security, speed, and workflow. OpenAI Codex operates differently. Engineers view it as a strict security architect. It generates highly secure, conservative scaffolding, such as signed OAuth states, but it struggles with creative frontend autonomy and operates slowly.
Google Antigravity 2.0 excels as a rapid multi-tool orchestrator for ambitious prototyping. You can build fast. The tradeoff is that users frequently report UX bugs and crashes during the actual execution phase.
Developers expecting a single tool to handle secure backend architecture and flawless frontend design are out of luck. The current state of AI coding requires assigning the right model to the right task.