Simon Willison gave the exact same one-shot game prompt to Claude Fable 5 and Codex Desktop running GPT-5.6 Sol Ultra. Codex produced the better game: Moonlight & Mayhem, a small 3D museum heist where you rescue two raccoon crewmates, stack the trio, steal the Golden Sardine, and escape.

The comparison is a useful agentic-game-development test:

  • Codex understood the premise better. Claude’s version reduced the crew to one raccoon collecting coins and fish in a backyard; Codex preserved the team-heist fantasy.
  • It generated and committed its own gpt-image-2 textures, with the prompts and outputs available in the repository.
  • The run took 52 minutes and used aggressive sub-agents. AgentsView estimates the equivalent API cost at $23.28: 700.7K input tokens, 32.5M cached tokens, and 148K output tokens.
  • Simon published the complete Codex transcript, making the whole build inspectable rather than presenting only the polished result.

The failure is as instructive as the win. Every raccoon shipped with an eyeball enlarged into a giant floating black sphere. Codex reviewed screenshots during development but never recognized the obvious visual defect. Simon fixed it with two prompts: “Why do the raccoons have huge black spheres on them?” and “Fix it.”

That is the practical lesson: current coding agents can independently turn a loose premise into a coherent, playable browser game — including mechanics, assets, mobile controls, and deployment — but screenshot access is not the same thing as reliable visual QA. Human taste and a final playthrough still matter.