A Breakout game that plays itself: Claude Fable 5.1 vs GPT-6 Astra

One brief to two AI agents: build a Breakout game in a single HTML file that plays itself. The paddle has to track the ball on its own from the moment the page loads, with a score and lives, and no click needed to start or keep going.

Claude Fable 5.1 ran in Claude Code, GPT-6 Astra in Codex CLI. One attempt each, no retries, no cherry picking, a fresh empty folder per run. Both shipped a game that runs and plays itself. Fable built a tall portrait board, Astra a wide editorial one.

The costs need a correction. The reel shows $0.77 against $0.47. My script at the time under-counted every Claude Code run, and Fable's real bill, read off its own transcript, was $2.27. Astra's $0.47 was right. So the gap was about 4.8 times, not 1.6. I left the reel up as posted and fixed the script for every round after it.

The fairness caveat still applies. The two ran in different harnesses, and Claude Code wraps the model in far more instructions than Codex does before my brief even arrives. That lands on the bill. So these are what each run really cost, not a clean model against model price.