110 Millisecond Vorbis-encoded audio file hits 4 million(!) input tokens

This audio file is extremely short, it’s basically just an utterance. This bug only arose once we switched from MP3 to VORBIS encoding for audio parts.

+How do we get refunded for the millions of nonsense tokens that are in our usage?

I’m posing here since it’s the link under “file a bug” in the Gemini API docs.

Since the file is not sensitive I am happy to share it if needed.

Hi @rubenkrupper,

Welcome to the Google AI Forum! :confetti_ball: :confetti_ball:

Thanks for clarifying about the file.. To help us reproduce the issue, please share the file.. Just DMed you..

@Krish_Varnakavi1 Where? Don’t see any DMs besides discobot:confused:

I don’t think you can get a refund, as AI isn’t promising you any good quality. A lot of this stuff is purely experimental and thus prone to errors and bugs.

1 Like

Hi @rubenkrupper,

Sent DM again.. Can you please check?