Vision Pricing calculator: inaccurate resizing/rounding vs API costs

API
, ,
1

Examples of online calculator’s issue vs API billings

gpt-image-1

Send an image to gpt-image-1 via the edits endpoint. The online calculator at the bottom of the pricing page shows:

This image shows a pricing and token calculation breakdown for processing an image of 1024 x 1025 pixels using the "gpt-image-1" model, resulting in a total price of $0.00323 for 323 tokens. (Captioned by AI)
This image shows a pricing and token calculation breakdown for processing an image of 1024 x 1025 pixels using the "gpt-image-1" model, resulting in a total price of $0.00323 for 323 tokens. (Captioned by AI)480×599 7.55 KB

However the API usage is:

Usage(input_tokens=219, input_tokens_details=UsageInputTokensDetails(image_tokens=194, text_tokens=25), output_tokens=272, total_tokens=491)

194 tokens = 1 tile is delivered by the API with an input image of 1024x1025, and also 1025x1024 (by a resize to 512x512 of the longer dimension, a round-down), not 323 = 2 tiles of the online calculator. (showing 1025->513).

Sending an actual 512x513 image gives two tiles’ billing.

Reducing to 1023x1024 input is one tile’s billing. I’ll let OpenAI run all the trials to find when “almost square” no longer becomes a square…it is formulaic - a different formula.

gpt-4.1-mini

One of these two calculators is wrong…

Mine

The image displays the OpenAI Vision Image Pricing Calculator interface showing input sizes, resized dimensions, token usage, billed tokens, and prices for two example images processed with the gpt-4.1-mini model. (Captioned by AI)
The image displays the OpenAI Vision Image Pricing Calculator interface showing input sizes, resized dimensions, token usage, billed tokens, and prices for two example images processed with the gpt-4.1-mini model. (Captioned by AI)699×408 16.9 KB

4000 x 1474 = 2450 billed tokens

OpenAI’s

The image shows two nearly identical pricing calculation panels for the "gpt-4.1-mini" model, with only a one-pixel difference in the set height (1474 px on the left, 1473 px on the right), resulting in a one-pixel difference in resized width but identical totals for tokens, patches, multiplier, and price. (Captioned by AI)
The image shows two nearly identical pricing calculation panels for the "gpt-4.1-mini" model, with only a one-pixel difference in the set height (1474 px on the left, 1473 px on the right), resulting in a one-pixel difference in resized width but identical totals for tokens, patches, multiplier, and price. (Captioned by AI)652×583 10.7 KB

4000 x 1474 = 2348 billed tokens ??

Trials

trial: send the image to gpt-4.1-mini

The image displays the multiplication equation "4000 x 1474" along with a text response noting it as a multiplication problem and offering to calculate the result. (Captioned by AI)
The image displays the multiplication equation "4000 x 1474" along with a text response noting it as a multiplication problem and offering to calculate the result. (Captioned by AI)433×482 16.2 KB

Showing total input 2498 billed tokens (w overhead)

The trick answer: Both are wrong, OpenAI worse

OpenAI found a way to bill the maximum

Billing is 1536 x 1.62 model multiplier = 2488

24 tiles × 64 tiles = 1536 token max

Additional

Edits endpoint, gpt-image-1

Beyond the prompt tokens measured, there are 32 additional text tokens that are billed when sending a single image. This is documented NOWHERE.

(postscript: my web script calculator obtained correct pricing when integrating gpt-image-1, why I had to triple-check a lot more cases and models.)