I originally said that GPT-5 was 10% better than Claude Sonnet 4. After using it with Cursor (gpt-5-fast) for a few days I've changed my mind, it's 20-30% worse. Firstly, it needs way too much handholding. I tell it to update a file with something and it will do half of it and then come back to me and ask if it should continue. I tell it "CONTINUE! DO NOT ASK ME ANY MORE QUESTIONS" and it will say "Absolutely! I am continuing without asking you any more questions.". Then it does a little bit and asks me whether it should continue again. I'm sure this is a good cost saver but not very beneficial to me.
Another problem it has is insisting that it's done things that it hasn't done yet. This is a problem that I haven't experienced with LLMs since the GPT-3 era. I say "update the wording in files X, Y and Z" - it will come back to me and say "I've successfully updated these three files and confirmed there are no linter errors [etc, etc]". I check the git status - no changes whatsoever! This seems to be a common issue and it's incredibly annoying. It seems convinced it's changed them, even if I tell it otherwise, so the solution is going back in the chat history and getting it to retry from a point.
Odd word choices are another one that feels like a regression way back to the past. A lot of the time I find that if I let it write something (an alt text, etc) it will come up with clipped sentences, words that don't make sense in the context, etc - maybe a side-effect of training on synthetic data? It at least doesn't have that flowery 4o sloppy feel to it but it doesn't feel like correct English - which is interesting because it's, again, not something I feel very much with modern models unless running a small quantised one locally, or at a high temperature.
I think that whatever they've done to make the model guess the user's intent from the prompt is bad. When viewing the reasoning it will often convince itself that I've asked it to do something else: it gives me an option and says it can continue if I ask. I say "continue" and it reasons into deciding that I really meant it should continue some other totally different thing.
Speed is an issue: it takes a while to begin reasoning and then once it's begun it takes another while to finish thinking. This isn't so bad normally but when you add 30s + 30s + 30s... with the model doing tool calls and thinking, it drags iteration time way down. The tokens/second speed is also poor in comparison to Sonnet 4. I didn't find this as much of an issue with the O series because it was much more effective at finding solutions, so I was happy to let it think about harder problems every now and then. But as a daily driver?
Coding quality doesn't feel that great either. It tends to just add and add - I let it go for a little bit on React and when I checked back it had written a 2800 line component that it took forever to handle, and wasn't very good at understanding itself. It also seems more likely than Claude to not fulfil all the user's instructions. I often give it a prompt with a few different things and it will only do some, totally giving up on the others until I prompt it again.
I have found it quite decent at writing SVGs and animations, so this might be somewhere it really shines. However I don't really write that many SVGs, and I do write a lot of code, so it's not that much of a boon to me.
All in all, 5/10. Definitely not feeling the AGI on this one, and for my next project I'll definitely be making Sonnet 4 the default model.
LinkedIn
Hacker
News
Github