Your reproduction on gpt-5.4-mini should be informed by:
“patches” being the way forward. You’ll have to resize smaller on images in your control if you want to pay less for less seeing.
You also will not see the embarrassing inflation in vision costs for comparable resolutions with the patches algorithm, multiplied by model token cost multipliers for vision tokens, multiplied by tripling input token pricing, in the pricing page vision calculator - it stopped being updated at gpt-5.2.
I will also note: gpt-5.4-mini is terrible. I tried to drop it into a gpt-5-mini app that rewrites a selection, following a prompt. gpt-5-mini follows the varying instruction, four paragraphs of output; gpt-5.4-mini is however basically non-functional with the same input, returning the app prompt or the original word back.
Unless I’m doing something extremely complex for vision-related data extraction, I still use gpt-5 with detail set to low. There are so many undocumented multipliers and unknown token consumption. They barely got the base token cap on low detail processing to work with gpt-5 after nearly a year of complaining on this forum.
Has anyone tested if this new patch algorithm for token consumption actually consumes the expected tokens?
No, nobody has ever tested?
Instead, nobody at OpenAI seems to fix billing issues after concrete reports in a timely manner.
Here is sending a 3000x2000 image to APIs today:
model
vision
vision_mult
chat input
calculated
responses input
calculated
gpt-5.5
patch
1.2
2699
2243
2888
2400
gpt-5.4
patch
1.2
2699
2243
2888
2400
gpt-5.2
patch
1.2
3051
2536
3310
2752
gpt-5.1
tile
-
917
910
917
910
Chat input: The vision request run on chat completions, usage report Responses input: Same request run on responses, usage report
Calculated: after removing the message overhead and reversing the cost multiplier, the number of image tokens that seem to be billed, per endpoint.
The cost multiplier is nowhere in documentation for vision for GPT-5+ full models.
We can see that GPT-5.2 specifically is incredibly over-billing today - a number that used to reconcile against OpenAI’s own calculator, where GPT-5.2 was the last model that they give the courtesy of a price calculator and what the billed tokens for that image should be:
GPT-5.4 and GPT-5.5, with a higher “high” token cap of 2500 dictating a larger resize, performs differently in resizing than the step-by-step documentation algorithm describes, and there is also different usage between endpoints for the same API call.
I have a calculator that previously was more accurate than OpenAI’s in providing token costs down to the token and subpixel of resizing done, even for indeterminate cases.
It implements the exact “patches” formulation described for downsizing (and also will show the uncapped pricing of detail:“original”)
What we expect - in resizing, in tokens, in inflation.
