Do you have a dumb question that you're kind of embarrassed to ask in the main thread? Is there something you're just not sure about?
This is your opportunity to ask questions. No question too simple or too silly.
Culture war topics are accepted, and proposals for a better intro post are appreciated.
Jump in the discussion.
No email address required.
Notes -
Do you find it reliably better than default 5? It seems to me that it's rather over-done and prone to skip ahead to something that is not necessarily what I want, rather than answering the specific query and working through with me as I prefer.
Yes, enormously so although "default 5" is also just not a high bar to clear (non-thinking 5 is similar quality to 4o, 5t is slightly better than o3 for most use cases other than "I want to run the 300 most obvious searches and combine the results in the obvious way in a table", where o3 still is unbeaten). 5T does seem to additionally be tuned to prioritize sounding smart over accuracy and pedagogy, and I haven't managed to tune the user instructions to fully fix this.
But yeah. Big difference.
More options
Context Copy link
I'm not a frequent enough LLM user to say how much of this was solid improvement vs luck, but my experience with free ChatGPT 5 (or any current free model, for that matter) versus paid GPT-5-Thinking was night vs day. In response to a somewhat obscure topology question, the free models all quickly spat out a false example (I'm guessing it was in the dataset as a true example for a different but similar-sounding question), and in the free tier the only difference between the better models and the worse models was that, when I pointed out the error in the example, the better models acknowledged it and gave me a different (but still false) example instead, while the worse models tried to gaslight me. GPT-5-Thinking took minutes to come back with an answer, but when it did the answer was actually correct, and accompanied by a link to a PDF of a paper from the 1980s that proved the answer on like page 6 out of 20.
I followed up with a harder question, and GPT-5-Thinking did something even more surprising to me: after a few minutes, it admitted it didn't know. It offered several suggestions for followup steps to try to figure out the answer, but it didn't hallucinate anything, didn't try to gaslight me about anything, didn't at all waste my time the way I'm used to my time being wasted when an LLM is wrong.
I've gotten used to using LLMs when their output is something that I can't answer quickly myself (else I'd answer it myself) but can verify quickly myself (else I can't trust their answer), but they seem to be on the cusp of being much more powerful than that. In an eschatological sense, maybe there's still some major architectural improvement that's necessary for AGI but still eluding us. But in an economic sense, the hassle I've always had with LLMs is their somewhat low signal-to-noise ratio, and yet there's already so much signal there that all they really have to do to have a winning product is get rid of most of the noise.
More options
Context Copy link
If you know the right prompt, you can get the models to leak OAI's profile of you. That includes usage stats. I believe I'm now at 95%+ GPT-5T usage, and almost zero for plain 5. The only time I use it is by accident, when the app "forgets" that I chose 5T in the model picker.
For any problem where you need even a modicum of rigor, I can't see a scenario where I wouldn't pick 5T over 5. If I need an instant answer, I use Claude. The free tier lets you use 4.5 Sonnet without reasoning, but it's still solid.
I will admit that I have barely used 5, because I gave it a few tries, found it barely better than 4o, and never touched it again. I just like 5T too. It has a bit of o3 in it, even if not quite as autistic. I really appreciate the lack of nonsense or sycophancy. 5 is far from the Pareto frontier on any aspect I care about.
More options
Context Copy link
More options
Context Copy link