Worth adding for anyone budgeting from this: fewer generated tokens per agent task doesn't always mean lower cost, because in multi-step runs the input side (the growing context resent every step) often costs more than the output. Before switching, I'd measure input tokens, output tokens and number of steps on a few of your real workloads. Also check the task success rate, since a shorter run that fails more often ends up costing more.
Yunus Emre Vurgun
Worth adding for anyone budgeting from this: fewer generated tokens per agent task doesn't always mean lower cost, because in multi-step runs the input side (the growing context resent every step) often costs more than the output. Before switching, I'd measure input tokens, output tokens and number of steps on a few of your real workloads. Also check the task success rate, since a shorter run that fails more often ends up costing more.