GPT-5.6 Sol Ultra, running through Codex Desktop with aggressive sub-agent usage, built a more faithful raccoon heist game than Claude Fable 5 from the same single prompt. Simon Willison ran the identical input through both systems. Fable 5 produced a lone raccoon collecting coins in a backyard. Sol Ultra produced a museum heist where you rescue two crewmates, stack them on top of each other, and steal a golden sardine from its case. The game, Moonlight and Mayhem, is playable now and the full GitHub repository is public.
The session cost an estimated $23.28 at full API prices: 700.7K input tokens, 32.5M cached tokens, 148K output tokens, across 52 minutes of agent runtime. Sol Ultra also generated its own textures using gpt-image-2. Despite reviewing screenshots during development, Codex shipped a build where each raccoon had a giant black polygon sphere floating over its head in place of an eye. Willison fixed it with two prompts: 'Why do the raccoons have huge black spheres on them?' followed by 'Fix it.'
The full Codex transcript is in the repository, and that is where this piece earns a full read. The transcript shows exactly how an aggressive multi-agent coding session moves, where it stalls, and what it misses even when given visual feedback. Willison also notes that Codex has a 'copy as Markdown' feature he wishes Claude Code would match. The gap between what the agent produced autonomously and what required a two-word human correction is the real story here.
[READ ORIGINAL →]