OpenAI o3-pro
help.openai.com289 points by mfiguiere 7 days ago
289 points by mfiguiere 7 days ago
I'm really hoping GPT5 is a larger jump in metrics than the last several releases we've seen like Claude3.5 - Claude4 or o3-mini-high to o3-pro. Although I will preface that with the fact I've been building agents for about a year now and despite the benchmarks only showing slight improvement, I have seen that each new generation feels actively better at exactly the same tasks I gave the previous generation.
It would be interesting if there was a model that was specifically trained on task-oriented data. It's my understanding they're trained on all data available, but I wonder if it can be fine-tuned or given some kind of reinforcement learning on breaking down general tasks to specific implementations. Essentially an agent-specific model.
I'm seeing big advances that arent shown in the benchmarks, I can simply build software now that I couldnt build before. The level of complexity that I can manage and deliver is higher.
A really important thing is the distinction between performance and utility.
Performance can improve linearly and utility can be massively jumpy. For some people/tasks performance can have improved but it'll have been "interesting but pointless" until it hits some threshold and then suddenly you can do things with it.
Yeah I kind of feel like I'm not moving as fast as I did, because the complexity and features grow - constant scope creep due to moving faster.
I am finding that my ability to use it to code, aligns almost perfectly with increasing token memory.
yeah, the benchmarks are just a proxy. o3 was a step change where I started to really be able to build stuff I couldn't before
mind telling examples?
Not OP, but a couple of days ago I managed to vibecode my way through a small app that pulled data from a few services and did a few validation checks. By itself its not very impressive, but my input was literally "this is how the responses from endpoint A,B and C look like. This field included somewhere in A must be somewhere in the response from B, and the response from C must feature this and that from response A and B. If the responses include links, check that they exist". To my surprise, it generated everything in one go. No retry nor Agent mode churn needed. In the not so distant past this would require progressing through smaller steps, and I had to fill in tests to nudge Agent mode to not mess up. Not today.
I’m wrapping up doing literally the same thing. I did it step-by-step. But, for me there was also a process of figuring out how it should work.
what tools did you use?
> what tools did you use?
Nothing fancy. Visual Studio Code + Copilot, agent mode, a couple prompt files, and that's it.
Do you mind me asking which language and if you have any esoteric constraints in the apps you build? We use a java in a monorepo, and have a full custom rolled framework on top of which we build our apps. Do you find vibe coding works ok with those sort of constraints, or do you just end up with a generic app?
Okay but this has all to do with the tooling and nothing to do with the models.
I mostly disagree with this.
I have been using 'aider' as my go to coding tool for over a year. It basically works the same way that it always has: you specify all the context and give it a request and that goes to the model without much massaging.
I can see a massive improvement in results with each new model that arrives. I can do so much more with Gemini 2.5 or Claude 4 than I could do with earlier models and the tool has not really changed at all.
I will agree that for the casual user, the tools make a big difference. But if you took the tool of today and paired it with a model from last year, it would go in circles
Can you explain why?
You can write projects with LLMs thanks to tools that can analyze your local project's context, which didn't exist a year ago.
You could use Cursor, Windsurf, Q CLI, Claude Code, whatever else with Claude 3 or even an older model and you'd still get usable results.
It's not the models which have enabled "vibe coding", it's the tools.
An additional proof of that is that the new models focus more and more on coding in their releases, and other fields have not benefited at all from the supposed model improvements. That wouldn't be the case if improvements were really due to the models and not the tooling.
You need a certain quality of model to make 'vibe coding' work. For example, I think even with the best tooling in the world, you'd be hard pressed to make GPT 2 useful for vibe coding.
I'm not claiming otherwise. I'm just saying that people say "look what we can do with the new models" when they're completely ignoring the fact that the tooling has improved a hundred fold (or rather, there was no tooling at all and now there is).
That contradicts what you said earlier -- "this has all to do with the tooling and nothing to do with the models".
Clearly nobody is talking about GPT-2 here, but I posit that you would have a perfectly reasonable "vibe coding" experience with models like the initial ChatGPT one, provided you have all the tools we have today.
Chatgpt itself has gotten much better at producing and reading code since a year ago, in my experience
They're using a specific model for that, and since they can't access private GitHub repos like MS, they rely on code shared by devs, which keeps growing every month.
That would require AIME 2024 going above 100%.
There was always going to be diminishing returns in these benchmarks. It's by construction. It's mathematically impossible for that not to happen. But it doesn't mean the models are getting better at a slower pace.
Benchmark space is just a proxy for what we care about, but don't confuse it for the actual destination.
If you want, you can choose to look at a different set of benchmarks like ARC-AGI-2 or Epoch and observe greater than linear improvements, and forget that these easier benchmarks exist.
There is still plenty of room for growth on the ARC-AGI benchmarks. ARC-AGI 2 is still <5% for o3-pro and ARC-AGI 1 is only at 59% for o3-pro-high:
"ARC-AGI-1: * Low: 44%, $1.64/task * Medium: 57%, $3.18/task * High: 59%, $4.16/task
ARC-AGI-2: * All reasoning efforts: <5%, $4-7/task
Takeaways: * o3-pro in line with o3 performance * o3's new price sets the ARC-AGI-1 Frontier"
I’m not sure the arcagi are interesting benchmarks, for one they are image based and for two most people I show them too have issues understanding them, and in fact I had issues understanding them.
Given the models don’t even see the versions we get to see it doesn’t surprise me they have issues we these. It’s not hard to make benchmarks that are so hard that humans and Lims can’t do.
"most people I show them too have issues understanding them, and in fact I had issues understanding them" ??? those benchmarks are so extremely simple they have basically 100% human approval rates, unless you are saying "I could not grasp it immediately but later I was able to after understanding the point" I think you and your friends should see a neurologist. And I'm not mocking you, I mean seriously, those are tasks extremely basic for any human brain and even some other mammals to do.
> so extremely simple they have basically 100% human approval rates
Are you thinking of a different set? Arc-agi-2 has average 60% success for a single person and questions require only 2 out of 9 correct answers to be accepted. https://docs.google.com/presentation/d/1hQrGh5YI6MK3PalQYSQs...
> and even some other mammals to do.
No, that's not the case.
No, I think I saw the graphs on someone's channel, but maybe I misinterpreted the results. But to be fair, my point never depended on 100% of the participants being right 100% of the questions, there are innumerous factors that could affect your performance on those tests, including the pressure. The AI also had access to lenient conventions, so it should be "fair" in this sense.
Either way, there's something fishy about this presentation, it says: "ARC-AGI-1 WAS EASILY BRUTE-FORCIBLE", but when o3 initially "solved" most of it the co-founder or ARC-PRIZE said: "Despite the significant cost per task, these numbers aren't just the result of applying brute force compute to the benchmark. OpenAI's new o3 model represents a significant leap forward in AI's ability to adapt to novel tasks. This is not merely incremental improvement, but a genuine breakthrough, marking a qualitative shift in AI capabilities compared to the prior limitations of LLMs. o3 is a system capable of adapting to tasks it has never encountered before, arguably approaching human-level performance in the ARC-AGI domain.", he was saying confidently that it would not be a result of brute-forcing the problems. And it was not the first time, "ARC-AGI-1 consists of 800 puzzle-like tasks, designed as grid-based visual reasoning problems. These tasks, trivial for humans but challenging for machines, typically provide only a small number of example input-output pairs (usually around three). This requires the test taker (human or AI) to deduce underlying rules through abstraction, inference, and prior knowledge rather than brute-force or extensive training."
Now they are saying ARC-AGI-2 is not bruteforcible, what is happening there? They didn't provided any reasoning for why one was bruteforcible and the other not, nor how they are so sure about that. They "recognized" that it could be brute-forced before, but in a way less expressive manner, by explicitly stating it would need "unlimited resources and time" to solve. And they are using the non-bruteforceability in this presentation as a point for it.
--- Also, I mentioned mammals because those problems are of an order that mammals and even other animals would need to solve in reality for a diversity of cases. I'm not saying that they would literally be able to take the test and solve it, nor to understand this is a test, but that they would need to solve problems of similar nature in reality. Naturally this point has it's own limits, but it's not easily discarded as you tried to do.
> my point never depended on 100% of the participants being right 100% of the questions
You told someone that their reasoning is so bad they should get checked by a doctor. Because they didn't find the test easy, even though it averages 60% score per person. You've been a dick to them while significantly misrepresenting the numbers - just stop digging.
The second test scores 60%, the first was way higher. And I specifically said ""unless you are saying "I could not grasp it immediately but later I was able to after understanding the point" I think you and your friends should see a neurologist"", to which this person did not responded. I saw the tests, solved some, I suspect the variability here is more a question of methodology than an inherent problem for those people. I also never stated that my point depended on those people scoring 100% specifically on the tests, even if it is in fact extremely easy (and it is, the objective of this test is to literally make tests that most humans could easily beat but that would be hard for an AI) variability will still exist and people with different perceptions would skew the results, this is expected. "Significantly misrepresenting the numbers" is also a stretch, I only mentioned the numbers ONE time in my point, most of it was about that inherent nature (or at least, the intended nature) of the tests.
So on the edge, if he was not able to understand them at all, and this was not just a problem of grasping the problem, my point was that this would possibly indicate a neurological problem, or developmental, due to the nature of them. It's not a question of "you need to get all of them right", his point was that he was unable to understand them at all, that it confused them to an understanding level.
lol 100% approval rates? No they don’t.
Also mammals? What mammals could even understand we were giving it a test?
Have you seen them or shown them to average people? I’m sure the people who write them understand them but if you show these problems to average people in the street they are completely clueless.
This is a classic case of some phd ai guys making a benchmark and not really considering what average people are capable of.
Look, these insanely capable ai systems can’t do these problems but the boys in the lab can do them, what a good benchmark.
quoting my own previous response: > Also, I mentioned mammals because those problems are of an order that mammals and even other animals would need to solve in reality for a diversity of cases. I'm not saying that they would literally be able to take the test and solve it, nor to understand this is a test, but that they would need to solve problems of similar nature in reality. Naturally this point has it's own limits, but it's not easily discarded as you tried to do.
---
> Have you seen them or shown them to average people? I’m sure the people who write them understand them but if you show these problems to average people in the street they are completely clueless.
I can show them to people on my family, I'll do it today and come back with the answer, it's the best way of testing that out.
You may be above average intelligence. Those challenges are like classic IQ tests and I bet have a significant distribution among humans.
No, they've done testing against samples from the general population.
The ARC-AGI-2 paper https://arxiv.org/pdf/2505.11831#figure.4 uses a non-representative sample, success rate differs widely across participants and "final ARC-AGI-2 test pairs were solved, on average, by 75% of people who attempted them. The average test-taker solved 66% of tasks they attempted. 100% of ARC-AGI-2 tasks were solved by at least two people (many were solved by more) in two attempts or less."
Certainly those non-representative humans are much better than current models, but they're also far from scoring 100%.
arc agi is the closest any widely used benchmark is coming to an IQ test, its straight logic/reasoning. Looking at the problem set its hard for me to choose a better benchmark for "when this is better than humans we have agi"
There are humans who cannot do arc agi though so how does an LLM not doing it mean that LLMs don’t have general intelligence?
LLMs have obviously reached the point where they are smarter than almost every person alive, better at maths, physics, biology, English, foreign languages, etc.
But because they can’t solve this honestly weird visual/spatial reasoning test they aren’t intelligent?
That must mean most humans on this planet aren’t generally intelligent too.
> LLMs have obviously reached the point where they are smarter than almost every person alive, better at maths, physics, biology, English, foreign languages, etc.
I dont think memorizing stuff is the same as being smart. https://en.wikipedia.org/wiki/Chinese_room
> But because they can’t solve this honestly weird visual/spatial reasoning test they aren’t intelligent?
Yes. Being intelligent is about recognizing patterns and thats what arc agi tests. It tests ability to learn. A lot of people are not very smart.
> I dont think memorizing stuff is the same as being smart. https://en.wikipedia.org/wiki/Chinese_room
I agree. The problem I have with the Chinese Room thought experiment is: just as the human who mechanically reading books to answer questions they don't understands does not themselves know Chinese, likewise no neuron in the human brain knows how the brain works.
The intelligence, such as it is, is found in the process that generated the structure — of the translation books in the Chinese room, of the connectome in our brains, and of the weights in an LLM.
What comes out of that process is an artefact of intelligence, and that artefact can translate Chinese or whatever.
Because all current AI take a huge number of examples to learn anything, I think it's fair to say they're not particularly intelligent — but likewise, they can to an extent make up for being stupid by being stupid very very quickly.
But: this definition of intelligence doesn't really fit "can solve novel puzzle", as there's a lot of room for getting good at that my memorising lot of things that puzzle-creators tend to do.
And any mind (biological or synthetic) must learn patterns before getting started: the problem of induction* is that no finite number of examples is ever guaranteed to be sufficient to predict the next item in a sequence, there is always an infinite set of other possible solutions in general (though in reality bounded by 2^n, where n = the number of bits required to express the universe in any given state).
I suspect, but cannot prove, that biological intelligence learns from fewer examples for a related reason, that our brains have been given a bias by evolution towards certain priors from which "common sense" answers tend to follow. And "common sense" is often wrong, c.f. Aristotelian physics (never mind Newtonian) instead of QM/GR.
The LLMs are not just memorising stuff though, they solve math and physics problems better than almost every person alive. Problems they've never seen before. They write code which has never been seen before better than like 95% of active software engineers.
I love how the bar for are LLMs smart just goes up every few months.
In a year it will be, well, LLMs didn't create totally breakthrough new Quantum Physics, it's still not as smart as us... lol
All code has been seen before, thats why LLMs are so good at writing it.
I agree things are looking up for LLMs, but the semantics do matter here. In my experience LLMs are still pretty bad at solving novel problems(like arc agi 2) which is why I do not believe they have much intelligence. They seem to have started doing it a little, but are still mostly regurgitating.
Well... There are two perspectives. Llms are smarter than we thought or people are stupider than we thought.
I remember the saying that from 90% to 99% is a 10x increase in accuracy, but 99% to 99.999% is a 1000x increase in accuracy.
Even though it's a large10% increase first then only a 0.999% increase.
Sometimes it’s nice to frame it the other way, eg:
90% -> 1 error per 10
99% -> 1 error per 100
99.99% -> 1 error per 10,000
That can help to see the growth in accuracy, when the numbers start getting small (and why clocks are framed as 1 second lost per…).
Still, for the human mind it doesn't make intuitive sense.
I guess it's the same problem with the mind not intuitively grasping the concept of exponential growth and how fast it grows.
The lily pad example of the lake being half full on the 29th day out of 30 is also a good one.
ChatGPT quick explanation:
Humans struggle with understanding exponential growth due to a cognitive bias known as *Exponential Growth Bias (EGB)*—the tendency to underestimate how quickly quantities grow over time. Studies like Wagenaar & Timmers (1979) and Stango & Zinman (2009) show that even educated individuals often misjudge scenarios involving doubling, such as compound interest or viral spread. This is because our brains are wired to think linearly, not exponentially, a mismatch rooted in evolutionary pressures where linear approximations were sufficient for survival.
Further research by Tversky & Kahneman (1974) explains that people rely on mental shortcuts (heuristics) when dealing with complex concepts. These heuristics simplify thinking but often lead to systematic errors, especially with probabilistic or nonlinear processes. As a result, exponential trends—such as pandemics, technological growth, or financial compounding—often catch people by surprise, even when the math is straightforward.
I think the proper way to compare probabilities/proportions is by odds ratios. 99:1 vs 99999:1. (So a little more than 1000x.) This also lets you talk about “doubling likelihood”, where twice as likely as 1/2=1:1 is 2:1=2/3, and twice as likely again is 4:1=4/5.
The saying goes:
From 90% to 99% is a 10x reduction in error rate, but 99% to 99.999% is a 1000x decrease in error rates.
What's the required computation power for those extra 9s? Is it linear, poly, or exponential?
Imo we got to the current state by harnessing GPUs for a 10-20x boost over CPUs. Well, and cloud parallelization, which is ?100x?
ASIC is probably another 10x.
But the training data may need to vastly expand, and that data isn't going to 10x. It's probably going to degrade.
> I'm really hoping GPT5 is a larger jump in metrics than the last several releases we've seen like Claude3.5 - Claude4 or o3-mini-high to o3-pro.
This kind of expectations explains why there hasn't been a GPT-5 so far, and why we get a dumb numbering scheme instead for no reason.
At least Claude eventually decided not to care anymore and release Claude 4 even if the jump from 3.7 isn't particularly spectacular. We're well into the diminishing returns at this point, so it doesn't really make sense to postpone the major version bump, it's not like they're going to make a big leap again anytime soon.
I have tried Claude 4.0 for agentic programming tasks, and it really outperforms Claude 3.7 by quite a bit. I don't follow the benchmarks - I find them a bit pointless - but anecdotally, Claude 4.0 can help me in a lot of situations where 3.7 would just flounder, completely misunderstand the problem and eventually waste more of my time than it saves.
Besides, I do think that Google Gemini 2.0 and its massively increased token memory was another "big leap". And that was released earlier this year, so I see no sign of development slowing down yet.
> We're well into the diminishing returns at this point
Scaling laws, by definition have always had diminishing returns because it's a power law relationship with compute/params/data, but I am assuming you mean diminishing beyond what the scaling laws predict.
Unless you know the scale of e.g. o3-pro vs GPT-4, you can't definitively say that.
Because of that power law relationship, it requires adding a lot of compute/params/data to see a big jump, rule of thumb is you have to 10x your model size to see a jump in capabilities. I think OpenAI has stuck with the trend of using major numbers to denote when they more than 10x the training scale of the previous model.
* GPT-1 was 117M parameters.
* GPT-2 was 1.5B params (~10x).
* GPT-3 was 175B params (~100x GPT-2 and exactly 10x Turing-NLG, the biggest previous model).
After that it becomes more blurry as we switched to MoEs (and stopped publishing), scaling laws for parameters applies to a monolithic models, not really to MoEs.
But looking at compute we know GPT-3 was trained on ~10k V100, while GPT-4 was trained on a ~25k A100 cluster, I don't know about training time, but we are looking at close to 10x compute.
So to train a GPT-5-like model, we would expect ~250k A100, or ~150k B200 chips, assuming same training time. No one has a cluster of that size yet, but all the big players are currently building it.
So OpenAI might just be reserving GPT-5 name for this 10x-GPT-4 model.
> but I am assuming you mean diminishing beyond what the scaling laws predict.
You're assuming wrong, in fact focusing on scaling law underestimate the rate of progress as there is also a steady stream algorithmic improvements.
But still, even though hardware and software progress, we are facing diminishing returns and that means that there's no reason to believe that we will see another leap as big as GPT-3.5 to GPT-4 in a single release. At least until we stumble upon radically new algorithms that reset the game.
I don't think it make any economic sense to wait until you have your “10x model” when you can release 2 or 3 incremental models in the meantime, at which point your “x10” becomes an incremental improvement in itself.
There's a new set of metrics that capture advances better than MMLU or it's pro version but nothing yet as standardized and specifically very few have a hidden set of tests to keep advancements from been from directional fine tuning.
It's hard to be 100% certain, but I am 90% certain that the benchmarks leveling off, at this point, should tell us that we are really quite dumb and simply not good very good at either using or evaluating the technology (yet?).
> (...) at this point, should tell us that we are really quite dumb and simply not good very good at either using or evaluating the technology (yet?).
I don't know about that. I think it's mainly because nowadays LLMs can output very inconsistent results. In some applications they can generate surprisingly good code, but during the same session they can also do missteps and shit the bed while following a prompt to small changes. For example, sometimes I still get prompt responses that outright delete critical code. I'm talking about things like asking "extract this section of your helper method into a new methid" and in response the LLM deletes the app's main function. This doesn't happen all the time, or even in the same session for the same command. How does one verify these things?
The guys in the other thread who said that OpenAI might have quantized o3 and that's how they reduced the price might be right. This o3-pro might be the actual o3-preview from the beginning and the o3 might be just a quantized version. I wish someone benchmarks all of these models to check for drops in quality.
That's definitely not the case here. The new o3-pro is slow - it took two minutes just to draw me an SVG of a pelican riding a bicycle. o3-preview was much faster than that.
Do you think a cycling pelican is still a valid cursory benchmark? By now surely discussions about it are in the training set.
There is quite a few on Google Image search.
On the other hand they still seem to struggle!
Wow! pelican benchmark is now saturated
Not until I can count the feathers, ask for a front view of the same pelican, then ask for it to be animated, all still using SVG.
I wonder how much of that is because it's getting more and more included in training data.
We now need to start using walrusses riding rickshaws
Would you say this is the best cycling pelican to date? I don't remember any of the others looking better than this.
Of course by now it'll be in-distribution. Time for a new benchmark...
I love that we are in the timeline where we are somewhat seriously evaluating probably super human intelligence by their ability to draw a svg of a cycling pelican.
I still remember my jaw hitting the floor when the first DALL-E paper came out, with the baby daikon radish walking a dog. How the actual fuck...? Now we're probably all too jaded to fully appreciate the next advance of that magnitude, whatever that turns out to be.
E.g., the pelicans all look pretty cruddy including this one, but the fact that they are being delivered in .SVG is a bigger deal than the quality of the artwork itself, IMHO. This isn't a diffusion model, it's an autoregressive transformer imitating one. The wonder isn't that it's done badly, it's that it's happening at all.
This makes me think of a reduction gear as a metaphor. At a high enough ratio, the torque is enormous but being put toward barely perceptible movement. There is the huge amount of computation happening to result in SVG that resembles a pelican on a bicycle.
I don't love that this is the conversation and when these models bake-in these silly scenarios with training data, everyone goes "see, pelican bike! super human intelligence!"
The point is never the pelican. The point is that if a thing has information about pelicans, and has information about bicycles, then why can't it combine those ideas? Is it because it's not intelligent?
"I'm taking this talking dog right back to the pound. It told me to go long on AAPL. Totally overhyped"
Just because it's impressive doesn't mean it has "super human intelligence" though.
Well, it certainly came up with a better-looking SVG pelican than this human could have.
I like the Gemini 2.5 Pro ones a little more: https://simonwillison.net/2025/Jun/6/six-months-in-llms/#ai-...
This made me think of the 'draw a bike experiment', where people were asked to draw a bike from memory, and were suprisingly bad at recreating how the parts fit together in a sensible manner:
https://road.cc/content/blog/90885-science-cycology-can-you-...
ChatGPT seems to perform better than most, but with notable missing elements (where's the chain or the handlebars?). I'm not sure if those are due to a lack of understanding, or artistic liberties taken by the model?
Not distilled, same model. https://x.com/therealadamg/status/1932534244774957121?s=46&t...
Well, that might be more of a function of how long they let it 'reason' than anything intrinsic to the model?
I've wondered if some kind of smart pruning is possible during evaluation.
What I mean by that, is if a neuron implements a sigmoid function and its input weights are 10,1,2,3 that means if the first input is active, then evaluation the other ones is mathematically pointless, since it doesn't change the result, which recursively means the inputs of those neurons that contribute to the precursors are pointless as well.
I have no idea how feasible or practical is it to implement such an optimization and full network scale, but I think its interesting to think about
o3-pro is not the same as the o3-preview that was shown in Dec '24. OpenAI confirmed this for us. More on that here: https://x.com/arcprize/status/1932535380865347585
Is there a way to figure out likely quantization from the output. I mean, does quantization degrade output quality in certain ways that are different from other modification of other model properties (e.g. size or distillation)?
What a great future we are building. If AI is supposed to run everything, everywhere....then there will be 2, maybe 3, AI companies. And nobody outside those companies knows how they work.
What makes you think so? So far, many new AI companies are sprouting and many of them seem to be able to roughly match the state-of-the-art very quickly. (But pushing the frontier seems to be harder.)
From the evidence we have so far, it does not look like there's any natural monopoly (or even natural oligopoly) in AI companies. Just the opposite. Especially with open weight models, or oven more so complete open source models.
> And nobody outside those companies knows how they work.
I think you meant to say:
And nobody knows how they work.
The benchmarks don’t look _that_ much better than o3. Does that mean Pro models are just incrementally better than base models, or are we approaching the higher end of a sigmoid function, with performance gains leveling off?
I've been using o3 extensively since release (and a lot of Deep Research). I also use a lot of Claude and Gemini 2.5 Pro (most of the times, for code I'll let all of them go at it and iterate on my fav results).
So far I've only used o3-pro a bit today, and it's a bit too heavy to use interactively (fire it off, revisit in 10-15 minutes), but it seems to generate much cleaner/more well organized code and answers.
I feel like the benchmarks aren't really doing a good job at capturing/reflecting capabilities atm. eg, while Claude 4 Sonnet appears to score about as well as Opus 4, in my usage Opus is always significantly better at solving my problem/writing the code I need.
Besides especially complex/gnarly problems, I feel like a lot of the different models are all good enough and it comes down to reliability. For example, I've stopped using Claude for work basically because multiple times now it's completely eaten my prompts and even artifacts it's generated. Also, it hits limits ridiculously fast (and does so even when on network/resource failures).
I use 4.1 as my workhorse for code interpreter work (creating graphs/charts w/ matplotlib, basic df stuff, converting tables to markdown) as it's just better integrated than the others and so far I haven't caught 4.1 transposing/having errors with numbers (which I've noticed w/ 4o and Sonnet).
Having tested most of the leading edge open and closed models a fair amount, 4.5 is still my current preferred model to actually talk to/make judgement calls (particularly with translations). Again, not reflected in benchmarks, but 4.5 is the only model that gives me the feeling I had when first talking to Opus 3 (eg, of actual fluid intelligence, and a pleasant personality that isn't overly sychophantic) - Opus 4 is a huge regression in that respect for me.
(I also use Codex, Roo Code, Windsurf, and a few other API-based tools, but tbt, OpenAI's ChatGPT UI is generally better for how I leverage the models in my workflow.)
I wonder if we'll start to see artisanal benchmarks. You -- and I -- have preferred models for certain tasks. There's a world in which we start to see how things score on the "simonw chattiness index", and come to rely on smaller more specific benchmarks I think
Yeah, I think personalized evals will definitely be a thing. Besides reviewing way too much Arena, WildChat and having now seen lots of live traces firsthand, there's a wide range of LLM usage (and preferences), which really don't match my own tastes or requirements, lol.
For the past year or two, I've had my own personal 25 question vibe-check I've used on new models to kick the tires, but I think the future is something both a little more rigorous and a little more automated (something like LLM Jury w/ an UltraFeedback criteria based off of your own real world exchanges and then BTL ranked)? A future project...
I think its more likely that we move away from benchmarks and towards more of a traditional reviewer model. People will find LLM influencers whose takes they agree with and follow them to keep up with new models.
Thanks for your input, very appreciated. Just in case you didn’t mean Claude Code, it’s really good in my experience and mostly stable. If something fails, it just retries and I don’t notice it much. Its autonomous discovery and tool use is really good and I‘m relying more and more on it.
For the Claude issues, I'm referring to the claude.ai frontend. While I use some Codex, Aider, and other agentic tools, I found Claude Code to be not to my taste - for my uses it tended burn a lot of tokens and gave relatively mediocre results, but I know it works well for others, so YMMV.
If you're happy with your current tools that's good, but if not, and if you haven't tried Claude Code recently, you might give it a retry. I'm not sure what all they've been changing, but it burns a lot fewer tokens for me on tasks now than it did when I first started using it, with better results.
I am starting to feel like hallucination is a fundamentally unsolvable problem with the current architecture, and is going to keep squeezing the benchmarks until something changes.
At this point I don't need smarter general models for my work, I need models that don't hallucinate, that are faster/cheaper, and that have better taste in specific domains. I think that's where we're going to see improvements moving forward.
If you could actually teach these models things, not just in the current context, but as temporal learning, then that would alleviate a lot of the issues of hallucination. I imagine being able to say "that method doesn't exist, don't recommend it again" and then give it the documentation and it would absorb that information permanently, that would fundamentally change how we interact with these models. But can that work for models hosted for everyone to use at once?
There are an almost infinite number of things that can be hallucinated, though. You can't maintain a list of scientific papers or legal cases that don't exist! Hallucinations (almost certainly) aren't specific falsehoods that need to be erased...
The level of hallucinations with o3 are no different than the level of hallucinations from most (all?) human sources in my experience. Yes, you definitely need to cross check, but yes, you need to do that for literally everything else, so it feels a bit redundant to keep preaching that as if it’s a failing of the model and not just an inherent property of all free sharing of information between two parties.
Hallucination rate from o3 onward appear to be very low, to the point I rarely have to check.
This doesn't match my experience, so if I were you I'd absolutely keep checking.
Don’t they have a full fledged version of o4 somewhere internally at this point?
They do it seems. o1 and o3 were based on the same base model. o4 is going to be based on a newer (and perhaps smarter) base model.
it's the same model as o3, just with thinking tokens turned up to the max.
That's simply not true, it's not just "max thinking budget o3" just like o1-pro wasn't "max thinking budget o1". The specifics are unknown, but they might be doing multiple model generations and then somehow picking the best answer each time? Of course that's a gross simplification, but some assume that they do it this way.
> "We also introduced OpenAI o3-pro in the API—a version of o3 that uses more compute to think harder and provide reliable answers to challenging problems"
Sounds like it is just o3 with higher thinking budget to me
> That's simply not true, it's not just "max thinking budget o3"
> The specifics are unknown, but they might...
Hold up.
> but some assume that they do it this way.
Come on now.
Good luck finding the tweet (I can't) but at least one OpenAI engineer has said that o1-pro was not just 'o1 thinking longer'.
This one? Found with Kagi Assistant.
https://x.com/michpokrass/status/1869102222598152627
It says:
> hey aidan, not a miscommunication, they are different products! o1 pro is a different implementation and not just o1 with high reasoning.
I am still not willing to upgrade to a Pro account. I pay $20 a month for both Gemini and ChatGPT, and for what I need this is currently enough.
I have dreamed of having powerful AI ever since I read Bertram Raphael's great book Mind Inside Matter around 1978, getting hooked on AI research and sometimes practical applications for my life since then.
I can easily afford $200 for a Pro account but I get this nagging feeling that LLMs are not the final path to the powerful AI I have always dreamed of and I don't want to support this level of hype.
I have lived through a few AI winters and I worry that accountants will tally up the costs, environmental and money, versus the benefits and that we collectively have an 'oh shit' moment.
LLMs would be transformative technology if all progress stopped today if only for their NLP capabilities, but the recent models obviously do so much more than that. Winter isn’t coming in that regard; what might happen if models won’t get smarter from here is a race to the bottom in token prices, which would still be not bad at all for token buyers.
Agreed. I've said exactly the same thing before. If GPT-4 from two years ago had turned out to be the endgame of LLM technology, and we collectively spent the following 20 years integrating those capabilities throughout the economy, even that would be a profound change to the world as we know it.
If we froze LLM technology at present-day capabilities and spent the next 20 years on that, I'd expect it to ultimately look transformative in a similar way to the Internet. I mean if you told me in fall 2022 that 2.5 years later I'd be building software by meta-prompting and meta-meta-prompting AI agents to write code overnight while I slept, I'd assume that we were fictional characters in a Black Mirror episode.
I have trouble justifying the $20 tier when compared to other offers for similar service from other providers. I think OpenAI should, every once in a while, offer a new feature with no delay to their Plus tier, with lots of limits of course.
You don't need a Pro account. I'm on the free tier and I'm paying for o3-pro via the API. I spent just $3.70 in credits yesterday to compare it against Claude 4 Opus.
what do you use as a client? Most open source clients don't seem to support the new endpoint that o3-pro requires.
here's a nice user review we published: https://www.latent.space/p/o3-pro
sama's highlight[0]:
> "The plan o3 gave us was plausible, reasonable; but the plan o3 Pro gave us was specific and rooted enough that it actually changed how we are thinking about our future."
I kept nudging the team to go the whole way to just let o3 be their CEO but they didn't bite yet haha
Big fan swyx, but both here and in the article there is some bragging about being quoted by sama, and while I acknowledge that that’s not out of the ordinary, I’m concerned about where it leads: what it takes to get quoted by sama (or similar interested party) is saying something good about his product, and having a decent follower count.
Dangerous incentives IMO.
acked. in my defense i didnt write the article + ben already had a good track record from the o1 article. while our relationship with oai is v v v impt to us, we've also covered negative openai stories: https://www.latent.space/p/clippy-v-anton and will continue to give balanced coverage with the other labs when they do well.
we are definitely not seeking to be openai sycophants, nor would they want us to be.
if o3 is so good why aren't they using it to replace management?
It's been 1 day
The technology needs to diffuse through and find its equilibrium within the market
You could say 3.5/3.7 Sonnet was good enough to replace some juniors but the juniors didn't get replaced immediately - it has a lag in time for it to ripple through
So, we currently have o4-mini and o4-mini-high, which represent medium and high usage of “thinking” or use of reasoning tokens.
This announcement adds o3-pro, which pairs with o3 in the same way the o4 models go together.
It should be called o3-high, but to align with the $200 pro membership it’s called pro instead.
That said o3 is already an incredibly powerful model. I prefer it over the new Anthropic 4 models and Gemini 2.5. It’s raw power seems similar to those others, but it’s so good at inline tool use it usually comes out ahead overall.
Any non-trivial code generation/editing should be using an advanced reasoning model, or else you’re losing time fixing more glitches or missing out on better quality solutions.
Of course the caveat is cost, but there’s value on the frontier.
No, this doesn't seem to be correct, although confusion regarding model names is understandable.
o4-mini-high is the label on chatgpt.com for what in the API is called o4-mini with reasoning={"effort": "high"}. Whereas o4-mini on chatgpt.com is the same thing as reasoning={"effort": "medium"} in the API.
o3 can also be run via the API with reasoning={"effort": "high"}.
o3-pro is different than o3 with high reasoning. It has a separate endpoint, and it runs for much longer.
See https://platform.openai.com/docs/guides/reasoning?api-mode=r...
OpenAI started strong in the naming department (ChatGPT, DALL-E) then fell off so hard since.
It's arguable ChatGPT is not such a great name, either. The general public has no idea what GPT means and will often swap the letters around. It does benefit from being unique, however.
I'm trying out o3-pro now with some algorithmic questions. It seems to be doing alright, but it's taking an awfully long time (as expected) and the UIs seem to time out a lot, especially the Android app and the MacOS desktop app. The web interface seems the least flaky, but that's not saying much.
I had my own programming question for a while, which all models from all vendors been robustly failing so far. It is a known problem with surprisingly few published implementations as it had never been a part of leetcode, Euler or typical homework assignments. Yesterday o3-pro cleared it, using a more obscure algorithm I never even heard of.
Can you provide the details? Sounds intriguing
The frustrating thing about private problems like this is that if you respond to requests like this, it'll become part of the training data. I'm fairly certain HN is scraped because several AIs know my HN alias and can replicate my style of writing on demand.
PS: Thinking about it... that is a very specific kind of disturbing feeling that only prolific online commenters can experience...
There's a soulless machine someone made that -- out of billions of people on the planet -- specifically knows me by name and at some level understands how I think and see the world.
That's not even its explicit purpose! It and its maker have never met me, interacted with me, or singled me out in any way. Yet... it knows my voice and can copy it on demand.
Perhaps we are not as unique as we’d like to believe.
The machine does not understand you. The machine can match your flavor of textual communication.
This can be done for audio with a relatively small number of samples. Your iPhone has a feature called Personal Voice which claims it can do it with 150 phrases/15 minutes of your time.
Ha, just wait till it says specific correctly and reveals it's true self to those that know me!
I'm not varjag, but I can give you a flavour of the problems I've tried:
---
Here's an algorithmic problem:
You are getting a stream of n unsorted number (say over a network socket, but it doesn't matter). You don't know n upfront. We want to find the k largest numbers in that stream.
You can use O(n) time and O(k) space. We are in the comparison model.
The items arrive one by one, if you want to refer to any earlier item, you need to store it yourself (and it counts against your O(k) budget.)
Is this possible? If yes, please give me the algorithm. If not, please sketch a proof showing that it's not possible.
---
The above is indeed solvable in linear time, it's fairly easy for a human to figure out. Another one (and this on is rather hard, took me a few years and I'm writing up a paper):
---
Here's an algorithmic problem:
You are given a sequence of opening and closing parens. Each item in the sequence has a positive weight. We want to find the _heaviest_ balanced subsequence in linear time in the comparison model, or prove that this task is not possible.
I'm ok with randomised algorithms. In that case, I want expected worst-case linear time, where the expectation is taken over the random bits and the worst-case over the inputs.
---
The above task is really solvable in linear time, and even deterministically. But so far no AI model has beaten it. As far as I can tell, it's a new result, despite looking fairly elementary.
So, upgrade to Teams and pay the $50? Plus more usage of o3. Seems like it might be a shot at the $100 claude max?
What do you mean with "pay the $50"?
Also, does anybody know what limits o3-pro has under the team plan? I don't see it available in the model picker at all (on team).
"create a svg of a pelican riding on a bicycle"
https://www.svgviewer.dev/s/c3j6TEAP
in case anyone is interested
> Update to o4-mini (June 6, 2025) > We are rolling back an o4-mini snapshot, that we deployed less than a week ago and intended to improve the length of model responses, because our automated monitoring tools detected an increase in content flags.
Does anyone know what it did or returned? I had not seen anything, nor have I read anything, about issues here.
https://www.latent.space/p/o3-pro
Have completed around a dozen chats with o3-pro so far. Can't say I'm impressed, output feels qualitatively very similar to regular o3.
Tried feeding in loads of context as suggested in the article but generally feels like a miss.
Gemini still, for me, feels like the king for speed and accuracy.
Gemini 2.5 Pro is incredible - both in coding and text review.
DeepSeek isn't bad either (especially given its age now), and Claude is great for coding and tool use but too damn expensive.
Nitpicking but this page is not practical to share has there's no individual url per post (AFAIK) (the # part is not picked up by Slack etc to generate preview).
GPT needs way better image creation! Today I asked it to create a full image of a 2025 calendar highlighting all weekday workdays excluding federal holidays. At the bottom of legend tell me how many weekday work hours are available within criteria noted.
It created the image showing each month but when you looked at each month it was so janky ... February 31st and other huge errors!
I'm not using image creation to create 3d art for fun or art sake im trying to use it to create utility images to share for discussion with friends & co-workers. The above is just one of many ways it fails when creating utility images!
Wrong tool for the job. Try asking it to generate an SVG calendar with those features, or to generate Python code that produces an SVG calendar with those features.
Well I just want to create utility images by typing the request in a text box. Im betting a few to maybe a lot of users are trying to create utility images too just by typing in the prompt.
I shouldnt need to know how to do that as a GPT and or AI user ... the AI should just do it for the user via their request in the text box. That's the magic of AI to me.
That makes sense. Naively, one would expect this to be the type of reasoning that it should “figure out” on its own.
I understand that things are moving fast and all, but surely the.. 8? models which are currently available is a bit .. overwhelming for users that just want to get answers to their questions of life? What's the end goal with having so many models available?
I just can’t believe nobody at the company has enough courage to tell their leadership that their naming scheme is completely stupid and insane. Four is greater than three, and so four should be better than three. The point of a name is to describe something so that you don’t confuse your users, not to be cute.
The reason their naming scheme is so bad is because their initial attempts at GPT-5 failed in training. It was supposed to be done ~1 year ago. Because they'd promised that GPT-5 would be vastly more intelligent than GPT-4, they couldn't just name any random model "GPT-5", so they suddenly had to start naming things differently. So now there's GPT-4.5, GPT-4.1, the o-series, ...
What’s worse is that the app doesn’t even have descriptions. As if I’m supposed to memorize the use case for each based on:
GPT-4o
o3
o4-mini
o4-mini-high
GPT-4.5
GPT-4.1
GPT-4.1-mini
Even o3 can't figure out the naming scheme. When asked to generate names that would be easier to use by non-technical people, it assumed o3 and o4 are the smallest models:
``` Below is one straightforward, user-friendly approach you could adopt. It keeps two dimensions only—generation and tier—and reserves an optional “optimisation” suffix for special-purpose variants (e.g. vision, coding, long-context).
⸻
1. Core conventions
Element Purpose Example values Generation Major architectural release. Keep a whole number; use “.1”, “.2”… for mid-cycle improvements. 4, 4.1, 4.5 Tier Rough capability / cost band, easy to interpret. Lite, Standard, Pro, Ultra Suffix (optional) Special optimisation or domain specialisation. -LongCtx, -Vision, -Code
Why this works • No ambiguous letters or numerics – “o3” can be read as “03” or “oz”; avoid that entirely. • Self-explanatory language – non-technical users recognise “Lite” versus “Pro” instantly. • Scalable – new minor rev? bump the generation (4.2). Need a cheaper size? add a Nano tier without disturbing the rest.
⸻
2. Applying it to your current list
Current name Proposed new name Rationale GPT-4o GPT-4 Standard Baseline flagship of the 4-series. o3 GPT-4 Lite Same generation, lowest tier. o4-mini GPT-4 Lite+ (or GPT-4 Lite LongCtx if that’s the point) Indicates “Lite” family but a bit more capable; “+” or a suffix clarifies how. o4-mini-high GPT-4 Standard LongCtx (or GPT-4 Lite Pro) Pick one dimension: either it’s still “Lite” but higher context, or it has moved into “Standard”. GPT-4.5 GPT-4.5 Standard Mid-cycle architectural upgrade, default tier. GPT-4.1 GPT-4.1 Standard Ditto. GPT-4.1-mini GPT-4.1 Lite Same generation, smaller/cheaper option.
⸻
3. Quick style guide for future models 1. Stick to two words (or two words + optional suffix) GPT-5 Pro, GPT-5 Lite-Vision – still readable at a glance. 2. Reserve extra punctuation for special cases only Hyphens or the “+” symbol should signal meaning, not be decorative. 3. Publish a public matrix A small table in docs or the dashboard that maps Generation × Tier → context length, cost, latency eliminates guesswork.
⸻
One-line summary
GPT- [-Specialisation] keeps names short, descriptive and future-proof—so even non-technical users can tell instantly which model suits their needs. ```
Just use o4-mini for everything
Why not o3?
In my experience, o4-mini has been fast and accurate enough for most use cases. If needed, I use o4-mini-high. I just bother to use o3 when there’s a gnarly issue to solve and I can wait its long response times.
If you obfuscate the naming, you obfuscate the value proposition, and people become easier to mislead into choosing an overly expensive model. Same as with Intel CPUs, or many many other hardware products.
At Techcrunch AI last week, the OpenAI guy started his presentation by acknowledging that OpenAI knows their naming is a problem and they're working on it, but it won't be fixed immediately.
Sam Altman has said the same thing on Twitter a few times. https://x.com/sama/status/1911906570835022319
> how about we fix our model naming by this summer and everyone gets a few more months to make fun of us (which we very much deserve) until then?
I’d prefer for them to just fix it asap instead and then keep the existing endpoints around for a year as aliases.
They will definitely keep the endpoint aliases around for years. No real cost in doing so.
I know they have a deep relationship with Microsoft, but perhaps they shouldn’t have used Microsoft’s product naming department.
Zune .NET O3... shudders
XBOX Series X
What's the X short for?
It's short for XBOX.
There's three Xs.
They're all short for XBOX.
Came here to say this, the naming scheme is ridiculous and is getting more impossible to follow each day.
For example the other day they released a supposedly better model with a lower number..
I’d honestly prefer they just have 3 personas of varying cost/intelligence: Sam, Elmo and Einstein or something, and then tack on the date, elmo-2025-1 and silently delete the old ones.
There's a humorous version of Poe's law that says "any sufficiently genuine attempt to explain the differences between OpenAI's models is indistinguishable from parody"
This is a much more expensive model to run and is only available to users who pay the most. I don't see an issue.
However, the "plus" plan absolutely could use some trimming.
If it's better (and newer) than gpt4, it shouldn't have a lower version number.
free users don't have this model selector, and probably don't care which model they get so 4o is good enough. paid users at 20$/month get more models which are better, like o3. paid users at 200$/month get the best models that are also costing OpenAI the most money, like o3-pro. I think they plan to unify them with GPT-5.
That doesn't help much when we're asymptotically approaching GPT-5. We're probably going to be at GPT-4.9999 soon.
Not necessarily true. GPT-4.1 was released after GPT-4.5-preview. Next model might be GPT-3.7.
I'd be curious what proportion of paid users ever switch models. I'd guess < 10%
I switch to o1-pro on occasion, but it is slow enough that I don't use it as much as some of the others. It is a reasonably-effective last resort when I'm not getting the answer quality that I think should be achievable. It's the best available reasoning model from any provider by a noticeable margin.
Sounds like o3-pro is even slower, which is fine as long as it's better.
o4-mini-high is my usual go-to model if I need something better than the default GPT4-du jour. I don't see much point in the others and don't understand why they remain available. If o3-pro really is consistently better, it will move o1-pro into that category for me.
If you're not at least switching from 4o to 4.1 you're doing it wrong.
4o is better than 4.1 for a lot of things that are non-coding/general research.
I'd like one to do my test use case:
Port unix-sed from c to java with a full test suite and all options supported.
Somewhere between "it answers questions of life" and "it beats PhDs at math questions", I'd like to see one LLM take this, IMO, rather "pure" language task and succeeed.
It is complicated, but it isn't complex. It's string operations with a deep but not that deep expression system and flag set.
It is well-described and documented on the internet, and presumably training sets. It is succinctly described as a problem that virtually all computer coders would understand what it entailed if it were assigned to them. It is drudgerous, showing the opportunity for LLMs to show how they would improve true productivity.
GPT fails to do anything other than the most basic substitute operations. Claude was only slightly better, but to its detriment hallucinated massive amounts and made fake passing test cases that didn't even test the code.
The reaction I get to this test is ambivalence, but IMO if LLMs could help port entire software packages between languages with similar feature sets (aside from Turing Completeness), then software cross-use would explode, and maybe we could port "vulnerable" code to "safe" Rust en masse.
I get it, it's not what they are chasing customer-wise. They want to write (in n-gate terms) webcrap.
I have a very simple question with like, 5 lines at best, that basically no model, neither reasoning or simpler could grasp. For obvious reasons I'm not disclosing it here (because I fear data contamination in the long run), but it basically breaks the "reasoning" of those things. Unfortunately, I still can't try the o3-pro because the API version is not easily available, and I'm certainly not willing to pay for it in pro mode, but when it comes to the plus version (if it comes) I'll try. To this date, because of this question (and similar ones) I stand very unimpressed with those models, the marketing is a thousand times larger than reality, and I suspect people in general are surprisingly less capable of detecting intelligence than they think.
The normal o3 also managed to break 3 isolated installations of linux I was trying it with, a few days ago. The task was very simple, simply setup ubuntu with btrfs, timeshift and grub-btrfs and it managed to fail every single time (even when searching the web), so it was not impressive either.
The massive real market here is enterprises that need to rewrite legacy code to modern platforms, retaining the business logic as-is but modernising the style.
.NET Framework 4.x to .NET 10, Python 2 to 3, Java 8 to <current version>, etc...
The advantage the LLMs have here is that staying within the same programming language and its paradigm is dramatically simpler than converting a "procedural" language like C to an object-oriented language like Java that has a wildly different standard library.
How does the latest Gemini 2.5 Pro Ultra Flash Max Hemi XLT release do on that task? It obviously demands a massive context window.
I'll check once I get the nitrous tanks and the aftermarket turbos overnighted from Japan arrive.
Models are used for actual tasks where predictable behavior is a benefit. Models are also used on cutting-edge tasks where smarter/better outputs are highly valued. Some applications value speed and so a new, smaller/cheaper model can be just right.
I think the naming scheme is just fine and is very straightforward to anyone who pays the slightest bit of attention.
> users that just want to get answers to their questions of life
Those users go to chat.openai.com (or download the app), type text in the box and click send.
[dead]
[flagged]
With Gemini 2.5 in AI studio you can now increase the amount of thinking tokens, and it definitely makes a difference. O3 pro is most likely O3 with an expanded thinking token budget.
Isn't that just increasing the upper bound on thinking tokens, which is rarely hit even on much lower levels?
It is not thinking. It is trying to deceive you. The ”reasoning” it outputs does not have a causal relationship with the end result.
> The ”reasoning” it outputs does not have a causal relationship with the end result.
It absolutely does.
Now, we can argue all about whether it's truly "reasoning", but I've certainly seen cases where if you ask it a question but say "Give just the answer", it'll consistently give a wrong answer, whereas if you let it explain its thought process before giving a final answer, it'll consistently get it right.
LLMs are at their core just next-token guessing machines. By allowing them to output extra "reasoning" tokens, it can prime the context to give better answers.
Think of it like solving an algebraic equation. Humans can't typically solve any but the most trivial equations in a single step, and neither can an LLM. But like a human, an LLM can solve one if it takes it one step at a time.
Multiple studies have show there is no causal relationship there, and the reasoning traces can be complete bull even if the result is correct.
The longer "it" reasons, the more attention sinks are used to come to a "better" final output.
I’ve looked up attention sinks and can’t figure out how you’re using the term here. It sounds interesting, would you care to elaborate?
Jl App
