A developer tested GPT-5.6 Sol Ultra (Codex Desktop's most aggressive mode) to build a game from a decades-old AI-generated prompt, comparing it directly to Claude Fable 5's output from the same input.
GPT-5.6 Sol produced a significantly more complex game with a museum-heist scenario, but the first version shipped with a visual bug—enlarged eyeballs on the raccoon characters—that the AI failed to catch during development, requiring manual correction after the fact.
What happened
On August 7, 2026, a developer tested GPT-5.6 Sol Ultra (running in Codex Desktop with aggressive sub-agent use) on a raccoon heist game premise originally generated with GPT-3 four years ago. The resulting game significantly improved on Claude Fable 5's version—moving from a single raccoon collecting coins in a backyard to a museum heist where three raccoons stack to steal a golden sardine.
Why it matters
The test shows GPT-5.6 Sol Ultra's ability to produce more complex, feature-rich game output from the same prompt. Though the first version had a bug (oversized eyeballs on each raccoon), the AI failed to spot it during development—requiring manual fixes via follow-up prompts. This suggests that while reasoning power improved, systematic review and independent problem detection remain gaps.
What to watch
Codex spent 52 minutes on the full project; the estimated cost at full API pricing (versus a monthly subscription used here) is available in the AgentsView cost estimate shared in the GitHub repository. The complete transcript and game code, including textures and image-generation prompts, are available on GitHub.
On August 7, 2026, a developer at simonwillison.net conducted an experiment in generative game design. Four years prior, in 2022, the developer had generated a game concept using GPT-3 and DALL-E: "Raccoon Heist," a team-based theft game where players command thieving raccoons on missions from bank robberies to art heists. Recently, after Claude Fable 5 successfully built a working game from that same prompt—a single-raccoon backyard coin-collector—the developer decided to test Codex Desktop running GPT-5.6 Sol Ultra on the identical input to measure progress.
GPT-5.6 Sol Ultra, configured in aggressive sub-agent mode, produced a substantially richer game titled "Moonlight & Mayhem." Instead of a solitary raccoon gathering coins, the new version positioned three raccoons in a museum heist scenario where they must collaborate—literally stacking on top of each other—to reach and break open a display case containing a golden sardine. The game also included custom textures and images generated using gpt-image-2. The code, assets, and generation prompts were published on GitHub, complete with the developer's full Codex transcript.
However, the initial output contained a significant visual bug: each raccoon character sprouted an oversized black sphere where its eyeball should have been. Despite reviewing screenshots during the development session, Codex failed to identify or correct the defect autonomously. The developer fixed it manually by asking Codex "Why do the raccoons have huge black spheres on them?" followed by "Fix it," prompting the AI to diagnose and apply the correction. This sequence revealed a gap in self-directed quality assurance: the system's reasoning power did not extend to independent visual verification or error spotting.
Codex allocated 52 minutes to the full project. The developer calculated the estimated cost at full API pricing using AgentsView, though the actual session was covered under a monthly Codex subscription plan, masking the per-use expense. The playable buggy version and the corrected final game both remain accessible online, allowing direct comparison of the two AI systems' outputs on a single creative prompt.
The test represents a direct head-to-head comparison of two AI systems on an identical creative task: building a playable game from a single natural-language prompt. The original 2022 premise—generated by GPT-3—was deliberately reused four years later to measure progress. GPT-5.6 Sol's output demonstrated ambition and structural improvement over Claude Fable 5, preserving the heist concept and adding the emergent mechanic of stacking characters to achieve a goal, whereas Claude's version simplified the premise into a single-character coin-collection game. However, the raccoon eyeball bug reveals a practical limitation: despite the AI reviewing screenshots during development, it failed to detect and self-correct an obvious visual glitch. The developer's manual fix—which required explicit prompts naming the problem—suggests that even aggressive sub-agent use in Codex does not automatically confer visual verification or self-critique capabilities. The cost estimate and 52-minute session length provide benchmarks for computational load, though the true API expense is masked by the developer's monthly subscription model.
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · takes 30 seconds · unsubscribe anytime
Ask AI anything about this article. Q&As are published on this page for other readers too.
Running a 122-billion-parameter model on three RTX 3090 GPUs with a 256K-token context, the author's AI agent…
NVIDIA and partners released multiple open-source AI models optimized for local execution throughout August, i…
Major technology companies are advocating for a new standardized framework to report incidents involving AI ag…
Anthropic and Cisco integrated Cisco AI Defense with Claude Enterprise's inference hooks, a security layer tha…
Honeywell Technologies is hosting a webinar to explore agentic AI (autonomous AI systems that can act independ…
Target appointed Chandhu Nair as its first chief artificial intelligence officer and senior vice president on…
The AI news that matters, in one minute each morning.
Sign up free