4o did this too because the dataset is out of date. If you tell it Donald Trump is president now and was reelected it will generally search the web and correct itself. Sometimes you do have to insist to invoke the web search (and make sure web search is enabled, though iirc that’s default now). Not really a slam dunk, just them being cautious with updating current events into the model.

If you want to slam dunk gpt5’s intelligence just interact with it for a bit and ask it to do tasks. It’s pretty lackluster tbh. In my testing it out it really sucks at parsing language compared to 4o mini and competing models, a step backwards.

I tried to have it build a relatively simple pcb to drive a flashing led circuit. It kept getting confused because I would point out the schematics it spit out were piss poor and impossible to read because of basic shit like overlapping symbols and labels. Rather than fix that issue it started to erase necessary components and kept changing how everything was connected which fucked the logic. Eventually it just randomly introduced an arduino into the mix? This was just a basic led flasher with mosfet, diodes, and transistors, mcu and code was not necessary. I ran out of free tokens and gave up. 4o mini was able to do it (though while it couldn’t generate an image it described the connections correctly and made an ascii diagram that was correct)

Even just basic stuff: I asked it to translate an English phrase into Japanese and it did. I then asked it to translate a Japanese phrase into English by saying “translate (Japanese phrase)” and it did. I then asked it to translate an English phrase into Japanese by saying “translate (English phrase)” and it rewrote the English as English. This is a pretty basic one. While I didn’t explicitly write “translate (English phrase) into Japanese” it was obviously clear from context what I meant and at a minimum I wouldn’t want the phrase left in English given I asked for “translate”. Bad parsing.