Reasoning tokens hidden price question

API
, ,
1

Hello everyone, I ran into a billing problem, it says in my dashboard that I spent 1 million tokens, according to my calculations it should have been about ~ $5, I used GPT-5 models (mini/nano), but in the end my bill turned out to be $20

Has anyone ever encountered this? All the settings for the model were medium

2

Reasoning tokens count as Output Tokens.

1 Like
3

One of the faults of the usage page in the platform site: that token count on the right side? It only shows input tokens. Generated tokens are completely missing from that overview.

You will have to select a date range to investigate, and then drill down further, by endpoint, seeing model token counts for input and output, which then can only be correlated by daily costs in a “group by → line item” and then hovering over the day’s tally.

Ultimate transparency of token consumption is to record them yourself, pulling the usage out of every API call you make, through a proxy. Then you can see the usage delivered and overbilling.

1 Like
4

Yes, I understand, but it says in my panel that I spent 1,200,000 tokens, according to the roughest calculations I should have spent ~$7, but in fact this figure is 3 times more

5

Go to this particular usage page:

https://platform.openai.com/usage/chat-completions

While “chat completions”, it should also have calls made through Responses.

Be sure to clear any per-project filter at upper right next to the date range:

image

Then where you have “Input tokens” at top-left, it is completely non-obvious that you have a drop-down there, and additionally, that you need to check each type of bar graph that you wish to receive.

image
image300Ă—343 7.81 KB

You’ll have to make good selections, otherwise it could show the same thing twice. For example, input cached and input uncached add together to make input tokens, and you’d see double the usage (silly). So just “input tokens” and “output tokens” (reasoning is also output tokens).

Then for the period selected, you get days to hover over (UTC 00:00), and also the total tokens of the period:

image

You can then pick “group by model”, and finally drill down to the token counts. Here, just showing the output tokens checked in the drop-down, grouped by model, a small date range, finally you get a per-day hover that starts to deliver your output consumption.

image
image503Ă—226 2.96 KB

Then what is beyond frustrating: only that main bargraph will show output tokens. Regardless of your checked selections, the following usage graphs by models STILL only show input tokens.

Then you’ll have to apply the model costs yourself to that token count, and go across other surfaces in usage to make sure there are no alternate projects or endpoints separately reporting gpt-5 models. Tedious, probably on purpose.

So: Your billing is probably right, just needing correct interpretation, except for OpenAI:

  • not delivering a gpt-5 cache discount when the calls are designed to be cacheable,
  • billing for quality: high, for image input to gpt-5 when requesting “low”,
  • the playground and the “prompt preset” billing you to get nothing output from gpt-5 because of prompts storing a bad max_output_tokens value of 2048 instead of “unlimited”, cutting off while still within reasoning.
6

I figured out what the problem is, it’s all the fault of the GPT-Image, for some reason the tokens from this model are not displayed in the panel)

1 Like