Prior model prompt overbilling on “edits”

On prior models, OpenAI seems to be over-billing for input text on the edits endpoint

• On v1/images/generations, all models have the same 6 token overhead - your sent prompt as measured by tiktoken with o200k_base, plus 6 tokens (that is not documented, but at least uniform).

• On /v1/images/edits, however, that figure grows. There is a significant leak of the input image size into the input text tokens.

gpt-image-1.5 1024x1024 inputs

• Send 13 tokens of prompt with one 1024x1024 input, size:1024x1024: get billed for 28 tokens. Over by 22. Over expectation by 16.

• Send 13 tokens of prompt with one 1024x1024 input, size:1024x1536: get billed for 35 tokens. Over expectation by 16.

• Send 13 tokens of prompt with two 1024x1024 input, size:1024x1024: get billed for 50 tokens. Over expectation by 30.

gpt-image-1.5 1:3 ratio inputs

• Send 12 tokens of prompt with one 480x1408 input, size:1024x1536: get billed for 57 tokens.
• Send 13 tokens of prompt with two 480x1408 input, size:1024x1536: get billed for 94 tokens. Text token bill significantly exceeding actual prompt.

Nowhere is this explained and nowhere is it reasonable to send input images and get extra billing for text, differing by the size.

Image inputs to gpt-image-1.5 still use the “tiles” formula and pricing of gpt-image-1, (undocumented also), with the addition of its billed output.text_tokens for reasoning.