What's this about?
ARC's cost curve bends backwards: Astra at maximum reasoning effort scored 62.7% for $26K; at medium effort, 38.6% for $48K — better *and* nearly half the price.
What exactly explains it?
ARC-AGI-3 is interactive: each game action costs a turn. More thinking per move means fewer, better moves — the total bill is actions × cost-per-action, and the smarter run takes far fewer actions. Measure twice, cut once, at benchmark scale.
Why is this worth noting?
It breaks the standard assumption that reasoning depth is a cost dial. For agentic workloads where actions are expensive (API calls, real-world steps), maximum effort may be the *budget* option — a genuinely new line item in the model-selection calculus.