Comment by anthonypasq a day ago
they could keep the current model in chatGPT the same forver and 99% of users wouldnt know or care, and unless you think hardware isnt going to improve, the cost of that will basically decrease to 0.
all these models just use web search now to stay up to date. knowledge cutoffs arent as important. also fine tuning new data into the base model after the fact is way cheaper than having to retrain the whole thing from scratch
The cost of old models decreases a lot, but the cost of frontier models, what people use 99% of the time, is hardly decreasing. Plus, many of the best models rely on thinking or reasoning, which use 10-100x as many tokens for the same prompt. That doesn't work on a fixed cost monthly subscription.
im not sure that you read what i just said. Almost no one using chatgpt would care if they were still talking to gpt5 2 years from now. If compute per watt doubles in the next 2 years, then the cost of serving gpt5 just got cut in half. purely on the hardware side, not to mention we are getting better at making smaller models smarter.
I don't really believe that premise in a world with competition, and the strategy it supports -- let AI companies produce profit off of old models -- ignores the need for SOTA advancement and expansion by these very same companies.
In other words, yes GPT-X might work well enough for most people, but the newer demo for ShinyNewModelZ is going to pull customers of GPT-X's in regardless of both fulfilling the customer needs. There is a persistent need for advancement (or at least marketing that indicates as much) in order to have positive numbers at the end of the churn cycle.
I have major doubts that can be done without trying to push features or SOTA models, without just straight lying or deception.
People cared enough about GPT-5 not being 4o that OpenAI brought 4o back.
https://arstechnica.com/information-technology/2025/08/opena...
For programming it's okay, for maths it's almost okay. For things like stories and actually dealing with reality, the models aren't even close to okay.
I didn't understand how bad it was until this weekend when I sat down and tried GPT-5, first without the thinking mode and then with the thinking mode, and it misunderstood sentences, generated crazy things, lost track of everything-- completely beyond how bad I thought it could possibly be.
I've fiddled with stories because I saw that LLMs had trouble, but I did not understand that this was where we were in NLP. At first I couldn't even fully believe it because the things don't fail to follow instructions when you talk about programming.
This extends to analyzing discussions. It simply misunderstands what people say. If you try to do this kind of thing you will realise the degree to which these things are just sequence models, with no ability to think, with really short attention spans and no ability to operate in a context. I experimented with stories set in established contexts, and the model repeatedly generated things that were impossible in those contexts.
When you do this kind of thing their character as sequence models that do not really integrate things from different sequences becomes apparent.