The issue where the token usage fees for GPT-4.1 Mini and the actual amount billed are wildly inconsistent

[gami]

Could someone knowledgeable please explain this to me? When using the GPT API, the number of tokens shown in the logs and the actual amount being billed are quite different. Why do you think that is?

For GPT-4.1 mini (input: $0.4 / output: $1.6),
Input: 32,000 tokens = $0.0128
Output: 2,400 tokens = $0.00384

However, it clearly feels like several dollars are being consumed each time.

[_j]

I’m curious where the “feeling” arises?

In the usage on the platform site, you can go to “Responses and Chat Completions” by clicking above the little graphs, then “group by->model” and see your daily usage of that model, in input and output tokens.

Then back to the main usage, again for the same day period, and choose “group by->line item”, to evaluate only the cost of that model again, in dollars.

One place where you are billed more than you sent is in vision. This model uses “patches”, which can be up to 1536 tokens, easily, and then OpenAI multiplies that usage by 1.62x->2489: $0.001 per image, still small, but that can stack up when you run conversations again.

[gami]

Thank you for explaining. I checked it using the method you suggested, and I found a few interesting things.

First, it seems that gpt-4.1-mini, which appears in the logs, is not actually being used according to the usage data. Instead, the usage shows GPT-5.2 Codex.
So I suspect the model name shown in the logs is incorrect.

Also, the numbers still don’t add up.

I went through the logs, and roughly this is what it looks like:

For GPT-5.2 Codex (input: $1.75 / output: $14.0)

Input: 1,000 × 847 tokens = 847,000 tokens = $1.48225
Output: 26 × 847 tokens = 22,022 tokens = $0.308308

Total: $1.79

I also took a rough look at both input and output tokens. While 1,000 tokens is on the higher side, there are quite a lot of requests with around 700 tokens, or even 100 tokens.
So it’s fair to say that the token count above is actually on the high side.

But the billed amount is completely different, right?
What could be causing this?

[_j]

If you are making no use of the gpt-5.2-codex model…then somebody else is, apparently to the tune of $100/day.

Revoke all your API keys across all projects.

Find out where you are leaking API keys where others can discover and abuse your API account.

https://platform.openai.com/docs/guides/production-best-practices#api-keys

The use of models can be with conversations self-managed and “store”:“false”, which won’t show in logs of API calls.

[gami]

I’m using gpt-5.2-codex for coding, so I don’t think it’s wrong that I’m being billed for gpt-5.2-codex. I also think the gpt-4.1-mini entry shown in the logs is just a display bug.

What I can’t understand is that, as shown in the calculation above, even assuming the usage is entirely gpt-5.2-codex, the numbers don’t come close to adding up.

[_j]

Note: I see a $0 for gpt-4.1 mini on that day you show the calls, at the bottom of the graph in the legend where there are values shown. You getting billed less than the API calls you are making.

Are you enrolled in “Share inputs and outputs with OpenAI” for training, to get daily complementary tokens on chat models? That could be why you have the opposite of what you report - essentially no bill for gpt-4.1-mini on 2026-02-06.

https://platform.openai.com/settings/organization/data-controls/sharing

If you want your own retrieval of costs, in buckets by day, you can create an ADMIN key, to use the organization admin endpoints.

Here’s my own “getting started” script for you, I’ve shared both help for “costs” and “usage” retrieval (essentially helping other API users answer similar mysteries):

Prepaid $10 for gpt-image-1, generated lots of images via API, but usage + costs still show $0

You will first want to go into the usage page: https://platform.openai.com/settings/organization/usage Then near top, select “Group by” → “Line Item”. That should show a bar graph display for the date range with the billed parts by model usage when you hover over a bar, and a legend at the bottom with the total for the period selected. For your API call, maybe it is the “costs” endpoint not behaving like you expect. It is limited to 7 days by default, requiring pagination, not obeying a range…

I'm being charged for free tokens

Ensure you aren’t making a mistaken assumption about the quantity of “free” first.. The “complementary tokens” are 250,000 tokens, daily, on gpt-4.1 for lower payment tier organizations, where a day is measured from 00:00 UTC time. This is a “free” amount pooled with all other main models that are listed, not a free amount per-model. Then a larger token pool for mini models. 250k is a decent amount, but not a whole bunch of tokens, and it also considers input tokens in the accounting. It is po…

[gami]

@_j

Note: I see a $0 for gpt-4.1 mini on that day you show the calls, at the bottom of the graph in the legend where there are values shown. You getting billed less than the API calls you are making.

Are you enrolled in “Share inputs and outputs with OpenAI” for training, to get daily complementary tokens on chat models? That could be why you have the opposite of what you report - essentially no bill for gpt-4.1-mini on 2026-02-06.

https://platform.openai.com/settings/organization/data-controls/sharing

If you want your own retrieval of costs, in buckets by day, you can create an ADMIN key, to use the organization admin endpoints.

Here’s my own “getting started” script for you, I’ve shared both help for “costs” and “usage” retrieval (essentially helping other API users answer similar mysteries):

I’m not receiving any free tokens.

More importantly, as shown in the screenshot, the logs indicate that gpt-5.2-codex was never used at all.
In my own code, the model I explicitly specify for coding is gpt-5.2-codex. However, for some reason, the logs show that gpt-4.1-mini was used instead, while the usage/billing is charged to gpt-5.2-codex.

Looking at the log contents themselves, the commands I gave to the LLM clearly correspond to the entries labeled as gpt-4.1-mini.
Because of this, I believe that the gpt-4.1-mini label in the logs is a bug, and that the model actually used is gpt-5.2-codex.

And even assuming that gpt-5.2-codex is indeed being used, as mentioned earlier, the discrepancy in the billed amount is far too large.

[_j]

Note this quirk: the times shown in the log list items are unexpectedly normalized there to local time.

The API usage however is delineated at 00:00 UTC.

That means your true daily usage can go across days you read directly out of the listing without clicking further to see a call’s time offset.

This is how gpt-5.2-codex shows up:

Nothing is getting added to logs right now in a timely fashion, as expected with this flaky log service that takes constant “it’s broke again” prodding, so I cannot report on the quality of model ID reporting now. I called with gpt-5.2-codex and gpt-5.2 with deliberate “store”:“true”…to see no evidence.

You can see by my previous “store”, for posterity, retaining how much they violate the documentation’s “default 30 days” by it being 10 months ago, I avoid this.

More

hours later - still nothing new in Responses’ logs…