I had a work version of GPT do a very simple spreadsheet summary task for me yesterday. It took it 5 minutes to do it. I could probably have done it myself in 30 or so minutes. The heavily subsidised token cost of that task? 10 dollars. That's with a 10x subsidy. The actual compute cost was about 100 dollars. There's something seriously wrong there. It's going to crash and crash HARD.
if people think i'm lying or are just interested. The spreadsheet had 45 sheets. Each sheet had roughly 500 x 50 populated cells. Formatting was not exactly standard across all sheets. The prompt was something like "there is labelled column in each sheet, give me a simple list of all the items from all the sheets in that column and ignore duplicates." We can chose which model to use. The model I chose was one of the newer ones, I honestly can't remember which one, possibly GPT 5.5. It took 5 minutes or more to so and the stated cost for the task was 10 dollars, possibly even more. I can't recall the token amount.
EDIT: After looking around for a few hours I found an ijustvibecodedthis.com article that made it sliiightly cheaper to run (like 30% cheaper) but it is still completely overpriced
[link] [comments]