The published cost is only comparable when we know the unit, modality and conditions.
USD per million tokens when the source uses that unit.
Requests with more than 272,000 input tokens apply price multipliers; tools may add per-call charges.
Open profile and sources →Thinking is billed as output. Fast mode costs twice the base rate.
Open profile and sources →Introductory pricing until December 31, 2026; from 2027, standard rates double.
Open profile and sources →Includes retries, tooling, recovery, cache, and human review.
Account and region limits may differ from the theoretical capacity of the model.
The absence is retained as relevant information, not as a zero.
A low rate can end up being expensive if the workflow repeats work or requires extensive oversight.
Internal or thinking tokens are usually billed as output and depend on difficulty and the configured effort.
Search, maps, code execution, and other tools may have per-call pricing in addition to tokens.
Repeated prefixes and asynchronous processing reduce cost, but require designing the workflow to take advantage of them.
Measure retries, human validation, incidents, and recovery: they are part of the total operating cost.