GPT-5.6 Sol Ultra Builds Improved Raccoon Heist Game in 52 Minutes

On 7th August 2026, Simon Willison published a blog post detailing his experiment with Codex Desktop running GPT-5.6 Sol Ultra, an AI model that aggressively uses sub-agents. He posed the same prompt that had previously generated a simple 'Raccoon Heist' game with Claude Fable 5, but this time the result was a significantly more polished game titled 'Moonlight & Mayhem'. The game, available on GitHub, includes textures and prompts generated using gpt-image-2. Unlike Fable's version, which featured a single raccoon collecting coins in a backyard, GPT-5.6 Sol's version places the player in a museum, where they must rescue two raccoon crewmates and stack on top of each other to steal a golden sardine from its case. However, the one-shot prompt produced a notable bug: each raccoon had an enlarged eyeball floating over its head as a giant black sphere. Despite reviewing screenshots during development, Codex failed to spot and correct this issue. Willison shared the full Codex transcript in the repository, praising Codex's 'copy as Markdown' feature, which he wishes Claude Code had. The project took 52 minutes to complete, and Willison provided an AgentsView cost estimate for the session, noting it would have been more expensive at full API prices compared to his monthly Codex subscription. The post is a link post and includes a sponsor call for $10/month to receive a curated email digest of important LLM developments.