How to read this
Most agent budgets blow up in one of two places: sending a giant prompt on every request, or running a frontier model on steps a cheaper one handles fine. This calculator makes both visible. Drop your real per-request token counts in, then watch the comparison table — the gap between tiers is usually larger than people expect, and prompt caching often moves the needle more than switching models.
For the full reasoning, see AI agent cost math: when Haiku beats Sonnet and prompt caching without switching models.