AI's Improving Capabilities Are Not the Result of Falling Costs

According to METR's data, AI's time horizon is doubling every few months, suggesting that AI could soon automate many tasks or jobs. However, some worry that a sharp rise in inference cost will make automation impossible.

But this is a misreading of the data. The rise in inference cost is not because models have become more expensive, but because the task length that models perform has gotten longer. In other words, the cost ratio relative to the human labor that the model replaces has actually remained constant.

Analyzing the cost ratio — that is, 'the AI inference cost to solve a task' divided by 'the cost of a human performing the same task' — revealed three key facts:

• Across successive frontier models, the cost ratio at the 50% confidence time horizon has not increased.
• Among tasks the model successfully completed, longer tasks did not show a higher cost ratio than shorter tasks.
• Even when AI spending per task was limited to a tiny fraction of human cost, the upward trend in time horizon barely slowed.

In conclusion, investing additional compute to raise the cost ratio can extend the time horizon, but this only accelerates the current trend rather than being a factor that delays automation. Therefore, the timing of AI automation will not be delayed by cost issues, and will arrive quickly along the existing capability-improvement curve.