I kind of believe there is a bug for calculating the enqueued toekns for batch api.
Background: I sliced my data equally to 152 samples per slice and submit the batch one by one. I checked the average token usage (in+out) from finished batch is ~1050/sample. I tried it with a Tier 1 account and using GPT-4.1-mini.
First time it failed, I tried to resubmit that slice and it works. But when I submitted the next one, it always failed. It is very frustrating that you don’t know what is happening, and you have to wait 24h to continue.
That happens not the first time and I try to understand how to avoid it.