A few days ago, Simon Willison asked Claude Fable 5 to turn a four-year-old game idea — originally generated with GPT-3 and DALL-E — into a working game from a single prompt. “Raccoon Heist”: a raccoon running around a back yard collecting coins and fish. Nice, but modest. Now he’s handed the exact same prompt to Codex Desktop running GPT-5.6 Sol Ultra — the mode where Sol makes aggressive use of sub-agents. A head-to-head between two coding agents on the same task.
The better game — and the eyeball bug
Codex delivered. The result is called “Moonlight & Mayhem,” and it’s a lot more ambitious: you’re in a museum, rescuing your two raccoon crewmates, stacking on top of each other to bust the golden sardine out of its display case. An actual heist — not just coin collecting. Much more “heisty,” as Simon puts it.
There’s one catch: the pure one-shot version shipped with a glorious bug. Each raccoon had an enormous black spherical eyeball, four times the size of its body, floating over its head. And the best part — Codex reviewed screenshots during development and still didn’t catch it. Simon fixed it with two short prompts: “Why do the raccoons have huge black spheres on them?” and then “Fix it.”
The numbers, and a little envy
Codex spent 52 minutes on the project. AgentsView’s cost estimate: about $23.28 at full API prices — with 700,000 input tokens plus 32.5 million cached tokens and 148,000 output tokens. For a monthly-subscription user, essentially free.
One side note that will catch Claude Code users’ attention: Simon wishes Claude Code had Codex’s “copy as Markdown” feature for sharing whole transcripts cleanly.
My take
I like comparisons like this because they’re honest. Fable 5 built a small game fast and clean; Codex, with aggressive sub-agents, built a bigger and richer one — but with a bug it should have caught itself. That’s exactly where we are right now: the agents are getting startlingly good, but quality control still has gaps. If you work with Claude Code every day, don’t be shy about looking over the fence. Not because one tool “wins” — but because duels like this are the fastest way to learn where your own tools are strong and where they aren’t.
Sources: Simon Willison: Moonlight & Mayhem (Raccoon Heist by Codex + GPT-5.6 Sol Ultra)