Hi All,
I’m having an issue with the GPT-4 API where I believe I was overcharged. The usage report indicates that I’ve sent around 52 million tokens, but I know for certain that my usage was closer to 3 million tokens at most.
I’ve tried reaching out for help through the “help” function, but the support system seems to be a bot that keeps looping through irrelevant topics without providing a solution.
Has anyone else experienced this, or does anyone know how I can get proper assistance with this issue?
Thanks in advance for any guidance!
@Fezilian ,
From my experience, the tracking of what you send and receive via the API is very accurate.
The reason why I can drop such a statement?
I am keeping all my inference outputs for research purposes, and I cross-referenced several times to conclude that they were all correct…
Do you have track and trace data for yourself?
Perhaps it’s worth noting the API cost goes in two ways (sending and receiving tokens).
Another possibility is the use of GPTs (Agents). When conversations grow, the number of tokens can quickly become very large.
Thank you for your response. I agree that the API is generally accurate, which is why this issue caught me by surprise. I’ve run the same code with similar data before and used tiktoken to estimate the costs. While there is some variability between the estimated and actual costs, it usually stays just under $5. Because of this consistency, I didn’t track tokens closely, trusting that everything would work as expected.
Could you clarify whether using agents is a default behavior, or do I need to explicitly invoke it? I’m fairly certain that I haven’t been using agents 
@rik.doclo Thank you for your help! I started tracking and tracing the data myself and discovered that the API’s behavior had changed, causing it to return irrelevant text, which significantly increased the cost. I optimized my prompt to prevent this, and it resolved the issue.
It’s unfortunate that I lost a lot of money in the process, but I’m grateful for your assistance. At least now I can avoid similar issues in the future.
@Fezilian ,
Happy you were able to pinpoint the issue at hand.
Willing our not, it still comes down to well designed prompts.
P.S. We all pay learning money 
1 Like