← Lab
Bloons TD, 41 ways
multiple models · 2026-10-14
Every model got the same prompt and an empty folder: build a Bloons-style tower defense game, with nobody answering its questions. Claude Opus, Sonnet, Fable and Haiku and OpenAI's GPT-6 Astra, Sol, Luna and GPT-6.1 each ran at several effort levels, from low up to max. Six more builds came from teams, where Opus or Sonnet planned the game and agents running another model wrote all of the code. Kimi K3.1 Pro ran once, at extra-high, in Claude Code over OpenRouter.
Each build is scored out of 10 on graphics, complexity, quality and creativity. Some things worth trying:
- Opus 5.5 max scored highest, at 8.9. It's the only build with its own cast of monkeys.
- Sonnet 5.5 max cost $163 and took almost five hours, and it tied with Fable 5.1 low's 22-minute run.
- GPT-6.1 high is the best OpenAI build, and it cost $1.40.
- Opus plans, Haiku builds scores 2.3 points higher than Haiku 5.5 high on its own.
- Kimi K3.1 Pro scored 4.9 for $1.73, and named its game Canopy Command, the same name three GPT builds chose on their own.
Launch
Runs on goosefeathers-lab.web.app — a separate site, because these builds are AI-written code.