The balloon colour follows the bloon layers: under 5 5+ 6+ 7+ 8+
How it was run
- Same prompt, word for word, in an empty folder. Nobody answered the model's questions, and each build had to be its own work.
- Claude models ran in Claude Code, OpenAI models in Codex, and Kimi K3.1 Pro in Claude Code over OpenRouter. One run per model and effort level, except Sonnet 5.5 high, which ran twice.
- Teams ran in Claude Code: Opus 5.5 or Sonnet 5.5 planned and reviewed, and agents running another model wrote all of the code.
- Each build is scored 1 to 10 on graphics, complexity, quality and creativity against one written rubric. The overall score is the average of the four. Small gaps between scores don't mean much.
- The Haiku, GPT-6 Luna and team builds were scored from scripted playthroughs plus a code review. The rest were also played by hand.
- Costs are what each run would cost at the model's list API price, worked out from the tokens it used. The Claude and GPT runs actually used subscriptions; Kimi's was paid through OpenRouter. Times run from prompt to final report, except Kimi's, which leaves out a wait for a credit top-up.
- The builds are as the models left them, with test files and screenshots left out. Three of them (GPT-6 Astra high, GPT-6.1 high and the Opus + Luna team) loaded files from the site root, so those paths were made relative to let them run from a subfolder. Nothing else was changed.