Same prompt, three Claude models: all three tower-defense games run and play.

1 The prompt (identical for all three)

Make an empty folder, start Claude Code in it, and send:

Make a playable tower-defense game in a single index.html file (canvas, no libraries, no external files). A winding path, enemies that come in waves, click to place three kinds of towers (rapid fire, splash, slow) that cost gold, lives and gold shown on screen, and a demo mode where the game plays itself until a key is pressed. It should look great on a phone-sized screen.

2 The three games they built

Open each one, press a key to leave demo mode, pick a tower card and tap the grass to place it.

• Haiku 5.5 game
• Opus 5.5 game
• Fable 5.1 game

3 What each run cost

Claude Code 2.1.309, headless, medium effort, an empty folder per model, no plugins, hooks or MCP servers, started at the same moment. Cost is what Claude Code reported for the run (API-equivalent list price; our runs used a subscription).

Fable cost 41 times what Haiku did for this prompt. "Works" means: zero page errors in 24 s of demo mode, a key press leaves demo mode, and a placed tower spends gold, checked in headless Chrome at phone size.

4 Run it yourself

# one empty folder per model
mkdir haiku && cd haiku
claude --model claude-haiku-5-5 --effort medium
# then paste the prompt above and open index.html

# same for the others
claude --model claude-opus-5-5 --effort medium
claude --model claude-fable-5-1 --effort medium

List prices per million tokens (input / output): Haiku 5.5 $0.10 / $0.50, Opus 5.5 $4 / $20, Fable 5.1 $10 / $50 (Anthropic model overview).

5 Honest limits

One simple one-file spec, one run per model, no scoring or ranking of the three games. This says nothing about harder, longer or more open-ended work, where the bigger models are meant to earn their price. Start with the cheap model for a job like this and step up when the result is wrong, not by default.