Comparing GPT-5.6 Sol Ultra and Claude Fable 5 in Game Development

Simon Willison tested the game-building capabilities of GPT-5.6 Sol Ultra via Codex Desktop against Claude Fable 5 by prompting them to recreate a raccoon-themed heist game. GPT-5.6 Sol Ultra, utilizing aggressive sub-agents and gpt-image-2, generated a much more complex and thematic game called "Moonlight & Mayhem." This demonstration highlights the rapid advancement of LLM-driven coding agents in executing complex, multi-step tasks like game development from a single prompt. It showcases how agentic workflows with sub-agents can significantly improve output quality compared to standard models, albeit at a higher token cost. Although the generated game was superior, it initially included a bug where the raccoons had giant floating eyeball spheres, which the AI failed to detect via screenshots but fixed after a direct prompt. The 52-minute development session consumed over 700K input tokens and 148K output tokens, costing an estimated $23.28.