AI API prices are often quoted per million tokens. The unit helps comparison, but it does not describe the whole bill.
Separate input and output
Output can cost much more than input for the same model. Applications with long prompts and those with long answers therefore have different cost patterns. Compare services using your own request mix.
Check the conditions
Cached reads, images, tool calls and context length can carry separate rates. Check minimum deposits, balance expiry and refunds too.
Verify the charge
Record token usage and balance changes for a test request. If the quoted rate and actual charge disagree, check currency, multipliers and model mapping before continuing.