Which AI models write working Godot 4 code
Each model wrote GDScript for the same 15 Godot 4 tasks, and every script was run headless on Godot 4.7.2 and judged by what Godot printed. Tasks passed counts the tasks that passed on the first try or after one fix with Godot's output; first try counts those that needed no fix.
| Model | Maker | Run on | Tasks passed | First try | Cost of the run |
|---|---|---|---|---|---|
| Claude Opus 5.5 | Anthropic | 15 of 15 | 15 | – | |
| Claude Sonnet 5.5 | Anthropic | 15 of 15 | 15 | – | |
| GPT-6.1 Sol | OpenAI | 15 of 15 | 15 | – | |
| Claude Haiku 4.5 | Anthropic | 14 of 15 | 11 | – |
4 models, the newest run of each, every one made as a coding agent in Claude Code or Codex. Cost is shown where the run recorded what the provider charged.