Hi everyone! I’ve run into a weird issue. I used to work with the ChatGPT API (GPT-4), and the costs were totally fine. But recently, I’ve noticed that it’s charging me insanely high amounts – like 10 cents per single question
I haven’t changed anything in my setup, just regular calls to same as before. But now, every request seems to drain my balance like crazy It feels like something changed behind the scenes.
Has anyone else experienced this? How can I figure out what’s causing such high usage? I’d really appreciate any ideas or tips
Have you lookend into your activity (export data) to understand how many tokens are you using per request?
If you are using playground, make sure to start a new conversation otherwise it will charge for the whole thread each time.
Hey! First of all — huge thanks for replying, honestly, you’re faster than ChatGPT support itself That already made my day.
About the issue — yeah, I actually didn’t use the Playground at all, so the “thread charging” thing isn’t the case here. All my requests went directly via API, and I’ve been tracking usage pretty closely.
What’s weird is that nothing in my setup changed — but the number of requests slightly increased (maybe 2x), and even then, my total token usage per day never went over ~50,000 tokens (input + output combined). I exported all usage data to CSV and checked — everything seems reasonable there.
And yet, within a short period, my balance dropped by over $10, which is totally unexpected. I even switched to GPT-3.5 after that, and for just two short prompts I got charged $0.12 — which doesn’t add up based on the official pricing.
So either:
• the model switch didn’t actually happen (even though I specified gpt-3.5-turbo);
• or something changed internally with how billing is being calculated lately;
• or maybe some kind of caching is being ignored and it’s charging full price each time.
I also canceled my billing plan after that just to be safe (had ~$5 left). Now I’m wondering if those remaining funds will be restored if I restart billing again
Thanks again — really appreciate the help! Let me know if you’ve experienced anything like that or figured out a workaround.
You can obtain, store, and save the usage object returned with an API call that has token counts. Then audit that usage increase after giving the API some time to “rest” or by using a particular project and API key for that call that you can isolate.
Or you can, right at where the API call is being made, create a log of the calls.
You might discovery some buggy bug where you are looping, or find out that the model is calling your tool iterator multiple times.
GPT-4 is that expensive, though. You can use gpt-4o for typical chat.
I see. Well, one thing I can think of, would be to enable the store parameter when you make the request, so that you can verify on your logs how many tokens each specific request is using.
Since you haven’t been using playground, it would be interesting to try directly there to see if the cost stays the same as in your app, when using the specific model.
Either on playground or logs you should be able to see how many tokens you are using and multiply it by the costs.
It feels like something changed behind the scenes.
That already made my day.