Comment by fallmonkey

The estimation for output token is too low since one reasoning-enabled response can burn through thousands of output tokens. Also low for input tokens since in actual use there're many context (memory, agents.md, rules, etc) included nowadays.