google-gemini/generative-ai-android

The count of received tokens is many times higher than expected!

karloti opened this issue · 0 comments

Description of the bug:

Here I am describing a problem where when the received information from Gemini is about 2.1KB (in which half is official because of JSON and it can be said that it is about 1KB) I am charged 8KB tokens.

Which means that for one received character I spend 8 tokens
(1 character = 8 tokens).
Is this normal?

I will clarify that the symbols are in Cyrillic and the token cost may be higher. I still don't know exactly how they are calculated.
In the normal case, an average of 4 symbols consumes one token
(4 characters = 1 token)

On the other hand, I will say that I limited the candidates to only 1, but I found no difference when they were 4 by default.

image
image

Actual vs expected behavior:

I ended up getting a spend x30 tokens over expected.
Otherwise it looks like this.
If you expect to pay $100 at the end of the month, you will pay $3,000. I say that quite responsibly.

Any other information you'd like to share?

I sent an email with many details about this case to your colleague, but since I didn't get a reply, I thought I'd share my observations here.

I feel that using this Gemini 1.5 Pro has many risks. It should be used very carefully and one of the solutions is to have a credit in the billing as with other products.

image