Replay / ARC-AGI-1 · Workstreams · GPT-6 Luna, low

Demo-cache floor

The model alone, shown every demonstration it bought for the task again for free.

Share learned

+0.19 ±0.05

Stream, budget used

47.7 solved 60%

Epilogue, budget used

65.2 solved 43%

Model spend

$2.24 2,234 calls

Instances

300 10.3 hours

Budget used is the share of the cap of 16 that an instance was charged, averaged over the instances: a solved instance costs 1 for each wrong answer and each demonstration, an unsolved one all 16, so lower is better. The share learned reads the stream's budget against the learner itself reset, the same system wiped before every instance of the same stream: the share of the budget left to learn that the learner saves, 0 for a learner that keeps nothing and 1 for perfect memory.