OpenAI *must* document the input image pricing of gpt-image-2

Absolutely no documentation is provided that describes the billing received for input images of various sizes on the edits endpoint, when using gpt-image-2

Additionally, nothing describes the context and typical or maximum received image count from use of the image editing tool on the Responses API endpoint, which obviously runs a parallel context consuming model with double-billing and no usage report.

The costs billed in tokens do not relate in any way to the “patches” algorithm of gpt-image-1, where images are described as downscaled to 512px for the shortest dimension, and then the only thing that could increase would be the longest dimension, by 512px steps. Also, input_fidelity is not an API option accepted but described as “always on”, for a described additional 4k or 6k billed, whereas this is forced “high” on gpt-image-1.5, ignoring “low”, contrary to documentation in multiple places.

Instead, increasing the width, there are steps that are not monotonic:

width x height → image_tokens 511 x 512 → 1024 512 x 512 → 1024 delta=+0 513 x 512 → 1024 delta=+0 1023 x 512 → 512 delta=-512 1024 x 512 → 512 delta=+0 1025 x 512 → 528 delta=+16 1535 x 512 → 768 delta=+240 1536 x 512 → 768 delta=+0 1537 x 512 → 833 delta=+65 2047 x 512 → 1408 delta=+575 2048 x 512 → 1408 delta=+0 2049 x 512 → 1430 delta=+22

Then take increasing a square image, where it responds to smaller sizes, but then has a stalling point:

width x height → image_tokens 256 x 256 → 256 384 x 384 → 576 delta=+320 768 x 768 → 1024 delta=+448 1024 x 1024 → 1024 delta=+0 1536 x 1536 → 1521 delta=+497

Or then a 2:1 aspect ratio, where we initially suspect a pattern similar to before, but then that breaks down.

width x height → image_tokens 512 x 256 → 512 1536 x 768 → 1152 delta=+640 2048 x 1024 → 1458 delta=+306 4096 x 2048 → 1458 delta=+0

Requests were made at an output size of “1440x480”, quality:low. A ratio of 1521 tokens billed for input image to 54 tokens billed for output image. Increasing the quality did not change the costs, nor did output size.

Publish the algorithm and its steps for normalizing, resizing, and billing by the resulting input image.

Document clearly what the AI is “seeing” in terms of billed semantic patches.

100% accurate cost estimation of a request is required.