What's this about?
Simon Willison had GPT-5.6 Sol Ultra (in Codex Desktop) and Claude Fable each build his "Raccoon Heist" game, and reported Sol Ultra did an even better job than Claude — a concrete, hands-on head-to-head datapoint.
What is a "coding harness" like Codex Desktop?
The model is the brain; the harness is the body — the app that lets the model actually read your files, run code, and iterate (Codex Desktop for OpenAI, Claude Code for Anthropic). The same model can perform very differently in different harnesses, so "Sol beat Claude" is really "Sol-in-Codex beat Fable-in-Claude" on this task.
Why does it matter?
Independent, task-specific comparisons from a trusted practitioner are more useful than benchmark leaderboards — though one game build is an anecdote, not a verdict.