Toby Ord examines whether the costs of AI agents are rising exponentially alongside their increasing task capabilities, as measured by METR’s time-horizon benchmark. Analyzing METR data reveals that while AI models can handle progressively longer human-equivalent tasks, their hourly costs often rise sharply—sometimes approaching or exceeding human labor costs—suggesting that improvements in AI performance may come with unsustainable increases in compute expense. This indicates a growing divergence between what AI can achieve in principle and what is economically practical for real-world applications.
