Epoch AI estimates that since 2023, the lowest price required to achieve a fixed level of benchmark performance has fallen by roughly 13× per year. If that trend continues even partially, capabilities that are expensive enough to reserve for important tasks today could become cheap enough to run continuously and at enormous scale.
This matters because falling inference costs multiply everything else AI can do. A capable model that costs $100 to run is useful. The same capability at $10, then $1, then pennies changes what becomes economically possible. Companies can deploy thousands of agents (instead of dozens) and let them work for hours (instead of minutes), retry failed approaches, review one another's work, and attack problems in parallel. The important transition might be about making increasingly capable cognition abundant enough to use almost everywhere.