A Raccoon Heist Game Exposes How AI Coding Models Differ

Developer Simon Willison spent an afternoon comparing two frontier AI coding systems on the same task: building a browser game about raccoons stealing a golden sardine. The results, published in a blog post on August 7, 2026, offer a small but revealing window into how differently these models approach the same creative brief.

The premise itself had an odd history. Willison originally generated it four years earlier using GPT-3 and DALL-E, then resurrected it as a prompt for Claude Fable 5 and, separately, for Codex Desktop running GPT-5.6 Sol Ultra.

Claude Fable 5 played it safe. It produced a single-raccoon game in which the player scurries around a backyard collecting coins and fish, a modest but functional interpretation of the brief.

GPT-5.6 Sol Ultra went further. Its version, built inside Codex Desktop, placed players in a museum where they rescue two raccoon crewmates and physically stack on top of one another to reach and steal the golden sardine. That extra structural ambition tracks with how Sol Ultra operates: the model leans heavily on sub-agents, effectively delegating parts of the build to itself in parallel.

That complexity came with a cost. The initial Sol Ultra build shipped with a glaring visual bug: each raccoon rendered with a giant, floating eyeball sphere hovering above its head. Codex Desktop reviewed its own screenshots during development but never caught the defect.

Codex Desktop spent 52 minutes on the Raccoon Heist project and still missed an oversized eyeball floating above every raccoon.

Willison had to intervene manually, prompting Codex with direct questions about the enlarged spheres and asking for a fix, which resolved the issue. The episode underscores a persistent gap in current coding agents: the ability to generate visually complex, multi-agent-coordinated output does not guarantee the ability to self-verify what that output actually looks like.

The comparison is informal, a single prompt run once on each system, but it fits a broader pattern developers have flagged with agentic coding tools. More sub-agent delegation and architectural ambition can produce richer features, yet it doesn't automatically translate into better quality control. Screenshot review, in this case, was present but ineffective, suggesting that visual QA remains one of the weaker links in autonomous coding workflows even as the code-generation side keeps improving.

For teams evaluating Claude versus OpenAI's coding stack for creative or front-end-heavy projects, the takeaway is less about picking a winner and more about where oversight is still required. Ambitious output from sub-agent architectures may need a human pass specifically for visual correctness, not just functional testing.