I don't think that GPT-5 Pro is much better (if better at all) than o3-pro. It's markedly slower. Output quality is comparable. It's still quite gullible and misses the point sometimes. It does seem better, however slightly, at suggesting novel approaches to problem solving. My initial impressions are that 5-pro is maybe 0-2% more knowledgeable and 5-10% more inventive/original than o3-pro. The "tone" and character of the models feel exactly the same.

I'll agree that it's superhuman and state-of-the-art at certain tasks: Formal logic, data analysis, and basically short analytical tasks in general. It's better than any version of Grok or Gemini.

When it comes to writing prose, and generally functioning as a writing bot, it's a poor model, obviously and transparently worse than Kimi K2 and Deepseek R1. (It never ceases to amaze me that the best English prose stylists are the Chinese models. It's not just that they don't write in GPT's famous "AI style," it's to the point where Kimi is actually on par with most published poets.)

I think it is, I've been using these models for 6 hours a day for almost a year. At any given time I have 2 of the max subscriptions (right now grok and openai).

I have a bug that was a complex interaction between backend and front end over websockets. The day before I was banging my head against the wall with o3 pro and grok heavy, gpt5 solved it first try.

I think its also true that most people arent pushing the limits on the models, and dont even really know how to correctly push the limits. Which is also why openai is not focussed on the best models

Similar usage as me, but I don't see a difference between o3-pro and 5-pro. Sounds odd, but my impression is that o1-pro was better at creating complex independent small functions than o3-pro/5-pro.

Actually will agree that o1 pro was better than o3 at really deep bug finding/coding analysis. Which is also why i have the theory that they could just turn up the compute to show better results, but dont do to cost. O3 and GPT5 seem heavily quantized, o1 pro was more raw

Another thing I’ll add though, is o3 pro is better through the api than the chat website. They clearly constrain it unless you’re paying the absurd api cost

yes, I think that's right. o1-pro for sure was too expensive for them.

How are you using GPT5 for coding (Cursor?)

I use repo prompt to build the prompt manually with the code files. Then give it to ChatGPT in the console. Then I have cursor ingrate the results.

I actually think cursor alone is not that good

Has anyone noticed that OpenAI is cutting off submitted context in ChatGPT Pro?

If I send about 60k tokens, the model can't see the question (at the bottom of the text). I need to reduce it to 50k.

If I send two prompts with 40k tokens, the model can't see the beginning of the first prompt.

This seems quite unethical given they advertise 128k context, and I doubt it's an accident (since it runs in the direction of cost savings).

> Kimi is actually on par with most published poets

Could you provide some examples, please? I find this really exciting. I’ve never yet encountered an AI with good literary writing style.

And poetry is really hard, even for humans.

To each their own, but I find the idea of ai-generated poetry sad as hell. I simply can’t see poetry as a collection of evocative words judged without context, in a vacuum— is poetry not both an activity and a relationship to most people? A person deftly portraying some difficult-to-express facet of the human experience and just maybe it viscerally strikes a chord with other people? I just don’t understand how people don’t value the fundamental humanity of that process. Even prose. James Baldwin stories, word for word, would land a hell of a lot differently if they were written and published by Hemingway.

I 100% agree. I am inclined to think an AI may be able to develop a sense of what words carry in future -- they can analyse it -- but it still lacks real experience.

Plus their creative output in literary quality is dreary, dull, and dire. That's why I was so curious for the OP to share examples.

Maybe it's the exposure to Chinese? I've heard that training models on code first helps, so I could see it.

I've also heard hearsay that R1 is quite clever in Chinese, too.

Stylized prose is where Claude 3 Opus particularly shines due to its character training and multilingual performance. It's plagued with claudeisms and has a ton of other shortcomings, but it's still better than any current model at this, including K2, R1, and especially Claude 4. Too bad Anthropic basically reversed their direction on creative writing, despite reporting some improvements each time (which don't seem to be true in practice).

Agreed. Something else that might be driving this is that existing models essentially get the job done for most users, who —- unlike HN commenters (I promise this is human generated, em dash notwithstanding ;P) —- don't quite about the state of the art.

checked my network - no one is using GPT 5 Pro ...

any feedback is greatly appreciated!!! especially comparing with o3

I'm surprised how few people are using it

Are there tight rate limits to GPT-5 Pro or is it in practice uncapped as long as you're not abusive?

Is GPT-5 better than GPT-5 Pro for any tasks?

One of the things that I have realised is, at this moment in time, it's absolutely a bad idea to buy a subscription for any of the models right now.

The offerings are evolving and upgrading at quite a rapid pace, so locking into one company's offering, or another's, is really wasted money (Why pay 200/year upfront for something that looks like it will be outdated within the next month (or quarter))

> The real story is running these models at true performance max likely could go into the thousands per month per user.

A loss leader model like that failed for Uber, because there really wasn't any other constraints on competition doing the same, including under pricing to capture market share - meaning it's a race to the bottom plus a test on whose pockets were the deepest.

AI pricing is stuck on subscription rather than metering which means that it’s a race to the bottom. It’s not obvious how AI providers can change that as the service they offer is just a game to users who can, even reluctantly, switch off.

I THINK that they are going to have to change pretty soon, I would guess that they would either drop their free tier offerings OR find another way to pay for things (advertising maybe?)

I'm not enough of a Business Major to know how they could monetise things, but I am enough of a realist to think that they can't stay like this forever

> I THINK that they are going to have to change pretty soon, I would guess that they would either drop their free tier offerings OR find another way to pay for things (advertising maybe?)

I would be incredibly surprised if OpenAI didn't go for advertising, given that they hired Fidji Simo (who PM'd feed ads at Facebook, and introduced ads to Instacart).

The OpenAI pitch deck for product placement was leaked over a year ago now

Just pay per month then.

Even a monthly payment is unnecessary at this moment - again the evolution of the quality of the models, but also the free tier offerings for each model (you can still use the older model's free tier, it was working enough before)

I personally haven't tried GPT 5 yet, but I am getting all I need from Claude and Gemini.

Once I start experimenting with GPT 5.0 - I will still use Claude and Gemini when I run out of free uses.

Your argument was not to get into a long term commitment because then you might have to wait a year to be able to use the best model again...not that you could get by using a rotating cast of free tier AI models...

You are complaining that i have multiple options available to me for not paying for a subscription.

The reason for not buying into a monthly subscription is not a reason thati should put shouldn't buy a year long one and is not negated in any way

No one's mad at you. You just switched your arguments.

As for why pay for access when you can get on free tiers? Well, because the pay models are far far better than the free ones. That's it.

See, you keep making these accusations. Nothing got switched.

Your idea wasn't valid, no need for all the nastiness

The annoyance of having to switch to a different model and losing my context is enough for me to pay the 20 dollars a month.

These models make me much more productive anyway. That is worth far more than $20.

I've had Claude tell me that I've reached the maximum messages for a session, which was painful.

Having said that it was a good circuit breaker, Claude was stuck on three wrong solutions to an issue, and having to re explain it meant that I realised what the bug was without further input from Claude

edit: FTR where Claude was stuck (and where it helped me)

First, I have a TUI application, written in Go, that is using https://github.com/rivo/tview - I chose tview basically at random, having no background in any of the TUI libraries available, but tview looked like it was easy to understand.

I had written some of the code, but it wasn't doing what I wanted of it. I had gotten Gemini to help me, but it still wasn't doing what I wanted.

Claude first suggested that the application needed to be refactored to make it into an Event Driven MVC TUI app, which I absolutely agreed with.

Claude counselled me on my understanding of the Run() command, and how it was a blocking call (which I hadn't realised). The refactor/re-architecture fixed the way things were being run a great deal.

However I still had a bug that I could not fix, and Claude was equally stumped (as was Gemini)

When I clicked on any of the components in a flex row, the first component behaved as though it was the one that was clicked. When I clicked on any of the components in another flex row the last one behaved as though it was the one that had been clicked.

Claude repeatedly told me that the problem was that the closure that used the index of components (inside a loop) was the problem.

This was outright wrong because

1. That bug has been fixed in Go.

2. The code I was sharing with Claude had the "fix" of local := index - meaning that the local version should have been correct

3. I repeatedly showed Claude logs (that it suggested I create) that showed that ALL of the components were in fact receiving the click event.

The second solution that Claude was fixated on was that th components overlapped one another inside the flex box.

This was partially incorrect.

I told Claude that I could visually see that the components weren't overlapping.

I also told Claude that there were borders around each component that clearly weren't overlapping.

I gave Claude debug logs that showed that the mouse click co-ordinates were inside a single component.

The third issue that Claude claimed was the problem was a click propagation bug in the library.

Claude repeated these claims several times, despite me showing that it wasn't either of the first two things, and I could not find a bug/issue for the third.

The circuit breaking made me stop and think about things, and I realised that all I really had to do was say inside the box "If a click event has been received AND (and this was the fix) You (the component) now have focus then behave".

What I suspect is happening is that the flex box container receives the click, and then tells all of its children that a click happened.

What disappoints me is, if what I suspect is true, then I would have expected Claude to have known that either from the tview documentation OR from reading the tview code (Claude does seem well acquainted with the library)

If my suspicion is correct then the second and third issues Claude claimed were causing me the bug were partially true, that is the flex container is on top of the components, which is an overlap, just not the components themselves overlapping.

The second partial correctness is that the way that the click is being propagated to the components seems to be that the flex box container is telling all of its child components that a click occurred. This isn't really a bug, it's likely well documented (as if I would RTFM...) or clear from the code.

I don't understand how you can have the patience to deal with an LLM like that. As if you're dealing with a low skilled but highly stubborn intern. Sounds like an awful waste of time to me.

Ha!

I have been a SWE for about 15 years, the majority of my career has been dealing with people that have about the same understanding as the LLMs, who scream abuse at me because their suggested solutions were invalid (in fact you can see someone heading down that path in another thread on this page)

The LLMs are less likely to run to HR when I tell them to eff 0ff with their stupidity (tested - told Claude that it had already effing suggested the bad index one and it was wrong because..., to which it politely apologised.. and re-suggested another one of its three suggestions)

So, from that point of view - LLMs are superior to some of the "senior" developers I have the misfortune of having to deal with previously.

As for my patience - I don't think I am being patient so much as doggedly determined to finding what I know is a fixable bug, that is, I will just keep gnawing at what I think should be fixable until a solution comes along, or I find something else shiny to occupy my spare time with (this being two side projects - the code and the testing of how good LLMs really are)

If people are constantly running to HR in your midst then you might not be the mentor you think you are.

exhibit c - this is the guy that is also exhibit a

FTR, me being told what to do by the other individual hardly sounds like I'm the one mentoring... but failure to read and comprehend has already proved to be your style.

The last word.

Oh dear.

A click event being sent to all elements inside a container if it clearly occurred on just one is what I'd consider an event propagation bug. I've never seen that happen before. Even if it's documented, it's pretty strange.

exhibit b.

Yep, LLMs tend to get into weird snags like this, usually with either niche or state of the art stuff.

I tried to get a game prototype up and running with https://spacetimedb.com - GPT-5 was clearly working with outdated information and couldn't get anything done. It just reverted back to things that didn't exist (commands had changed, their arguments were different).

Same project, same GPT-5. This time it went in weird circles when I tried to use Deno as the backend. First it understood how to import Phaser to Deno, then it forgot. Then it remembered. Then it forgot. All within the same conversation with us trying to get SOMETHING on the browser =)

I can definitely understand that it feels like looking at early adopters of technology with hindsight. Calculators, home/office computers, media players.

BUT, of course, if you can get more value out of it inside a month or year than you put in, it doesn't really matter. Differences between frontier models feels pretty low to me currently, which was not always the case; even so, they're certainly far ahead of free models for my uses.

I did pay for OpenAI both in a $20/mo subscription and later for API tokens (came out cheaper than the subscription). Since Gemini 2.5 came out, though, I just abuse Google's free AI Studio (at least free in the US; idk what kind of geo-gating they do). I'm not paying Google, but they are keeping me from paying their competitors. It will take a large and hard-gated leap forward (or Google deciding my $150-250 worth of unpaid use has gone on long enough) to get me to pay again.

I agree here but also believe it was a way to expose better models to the masses. o3 was so spectacularly good. But a lot of people were still not using it. Even some of my friends who use ChatGPT daily I would say are you using o3 and get a blank stare.

So I think it’s also a way to push reasoning models to the masses. Which increases OpenAI’s cost.

But due to the routing layer definitely cost cutting for power users (most of HN)… except power users can learn to force it to use the reasoning model.

Honestly as a daily (pro) user of ChatGPT, I didn’t even know o3 was the best, I thought 4o incorporated it (hence the o)

I remember reading that 4o was the best general purpose one, and that o3 was only good for deeper stuff like deep research.

The crappy naming never helped.

Wait, you decided to pay $200 a month and you didn’t know that o3 was better than 4o?

Oops I pay $20/month. Not the $200 pro. With Claude, Pro is the $19/mo plan. And Max is the $100-$200/month plan. Naming confusion.

The poor context length of O3 really cripples it.

They are clearly building the product for mass adoption, not power users

I don't think Pro is usable via the API, otherwise I'd be testing it. Is it usable through Codex CLI, given they updated that to be able to use your subscription?

I don't think Codex CLI uses the Pro/Plus subscriptions, not yet at least.

[delayed]

Does Pro fix the fundamental issues he describes? I would think it would have to do that to "completely invalidate his opinion", rather than just be better than the base model.

He didn’t describe any fundamental issues.