GPT-5 Ushers In AI’s ‘Trough of Disillusionment’

There’s no doubt: OpenAI’s GPT-5 is a good model.

I’ve been talking to developers, who’ve had a chance to to use the AI now. As we predicted, the flagship AI improved in domains like coding—where it got better at both practical coding tasks and designing websites and apps—and creative writing. Developers told me that the new model is also better at ingesting and analyzing longer prompts, which makes it more useful for things like analyzing tons of healthcare data.

The answer to whether or not GPT-5 lived up to the hype, though, is less obvious.

Our readers already knew that the jump from GPT-4 to GPT-5 wouldn’t be as impressive as past model upgrades, thanks to our deep dive into the rocky path behind developing the latest model. Yet OpenAI executives kept on teasing GPT-5 throughout the week and building up hype on social media, culminating in an hour-long, highly-produced launch event reminiscent of the flashiest iPhone launches.

The truth is that the new model still has lots of shortcomings that users picked up on fairly quickly. For instance, several developers told me that GPT-5 has issues knowing when it should “think harder” (i.e. use more compute) on tougher problems, which it was supposed to be able to do automatically. That was a big selling point of the new model, as CEO Sam Altman had promised that combining the traditional GPT model line with the “o” reasoning models would make ChatGPT much simpler to use.