Cache read tokens

Cache read tokens are the tokens an agent reads back from the prompt cache instead of sending fresh. When a session re-uses the same context each turn, most of its input is cache reads, not new input, which is why a long session can be far cheaper than its raw token count suggests.

A cache read is billed at a fraction of the input price, roughly a tenth on most models and less on the newest ones. Put a run's four counts into the cost calculator and you can see the split for yourself: in a real run, cache reads often dominate the token count while barely moving the bill.