I wonder how much of the '5 release was about cutting costs vs making it outwardly better. I'm speculating that one reason they'd deprecate older models is because 5 materially cheaper to run?
Would have been better to just jack up the price on the others. For companies that extensively test the apps they're building (which should be everyone) swapping out a model is a lot of work.
> For companies that extensively test the apps they're building
Test meaning what? Observe whatever surprise comes out the first time you run something and then write it down, to check that the same thing comes out tomorrow and the day after.
reading all the shilling of Claude and GPT i see here often I feel like i'm being gaslighted
i've been using premium tiers of both for a long time and i really felt like they've been getting worse
especially Claude I find super frustrating and maddening, misunderstanding basic requests or taking liberties by making unrequested additions and changes
i really had this sense of enshittification, almost as if they are no longer trying to serve my requests but do something else instead like i'm victim of some kind of LLM a/b testing to see how far I can tolerate or how much mental load can be transferred back onto me
I've worked on many migrations of things from vX to vX + 1, and there's always a tension between maximum backwards-compatibility, supporting every theoretical existing use-case, and just "flipping the switch" to move everyone to the New Way. Even though I, personally, am a "max backwards-compatibility" guy, it can be refreshing when someone decides to rip off the bandaid and force everyone to use the new best practice. How exciting! Unfortunately, this usually results in accidentally eliminating some feature that turns out to be Actually Important, a fuss is made, and the sudden forced migration is reverted after all.
I think the best approach is to move people to the newest version by default, but make it possible to use old versions, and then monitor switching rates and figure out what key features the new system is missing.
These things have cost associated. In the case of AI models that cost comes in the form of massive amounts of GPU hardware. So, I can see the logic for OpenAI to not want a lot of users lingering on obsolete technology. It would be stupendously expensive to do that.
Probably what they'll do is get people on the new thing. And then push out a few releases to address some of the complaints.
>I think the best approach is to move people to the newest version by default, but make it possible to use old versions, and then monitor switching rates and figure out what key features the new system is missing.
See, one would think this would be the common sense approach and I thought was how they did it previously, no?
What's odd is that OpenAI didn't seem to feel it was worth doing this time around.
Taking away user choice is often done in the name of simplicity. But let's not forget that given 100 users, 60 are likely to answer with "no opinion" when asked what about their preference to ANY question. Does that mean the other 40% aren't valuable and their preferences not impactful to the other "we don't care" majority?
It's not totally surprising given the economics of LLM operation. LLMs, when idle, are much more resource-heavy than an idle web service. To achieve acceptable chat response latency, the models need to be already loaded in memory, and I doubt that these huge SotA models can go from cold start to inference in milliseconds or even seconds. OpenAI is incentivized to push as many users onto as few models as possible to manage the capacity and increase efficiency.
Unless the overall demand is doing massive sudden swings throughout the day between models, this effect should not matter; I would expect the number of wasted computers to be merely on the order of the number of models (so like, maybe 19 wasted computers) even if you have hundreds of thousands of computers operating.
Welcome to every OpenAI launch. Marketing page says one thing, your reality will almost certainly not match. It’s infuriating how they do rollouts (especially when the marketing page says “available now!” or similar but you don’t get access for days/weeks).
This doesn't seem to be the case for me. I have access to GPT-5 via chatgpt, and I can also use GPT-4o. All my chat history opens with the originally used model as well.
I'm not saying it's not happening - but perhaps the rollout didn't happen as expected.
This is not a deprecation and users still have access to 4o, in fact it's renamed to "gpt-5-main" and called out as the key model, and as the author said you can still use it via the API
What changed was you can't specify a specific model in the web-interface anymore, and the MOE pointer head is going to route you to the best model they think you need. Had the author addressed that point it would be salient.
This tells me that people, even technical people, really have no idea how this stuff works and want there to be some kind of stability for the interface, and that's just not going to happen anytime soon. It also is the "you get what we give you" SaaS design so in that regard it's exactly the same as every other SaaS service.
o3 was also an anomaly in terms of speed vs response quality and price vs performance. It used to be one of the fastest ways to do some basic web searches you would have done to get an answer if you used o3 pro you it would take 5x longer for not much better response.
So far I haven’t been impressed with GPT5 thinking but I can’t concretely say why yet. I am thinking of comparing the same prompt side by side between o3 and GPT5 thinking.
Also just from my first few hours with GPT5 Thinking I feel that it’s not as good at short prompts as o3 e.g instead of using a big xml or json prompt I would just type the shortest possible phrase for the task e.g “best gpu for home LLM inference vs cloud api.”
There must be a weird influence campaign going on.
"DEEP SEEK IS BETTER" lol.
GPT5 is incredible. Maybe it is at the level of Opus but I barely got to talk to Opus. I thought Opus was a huge jump from my limited interaction.
After about 4 hours with GPT5, I think it is completely insane. It is so smart.
For me, Opus and GPT5 are just other level. This is a jump from 3.5 to 4. I think more if anything.
I am not a software engineer and haven't tried it vibe coding yet but I am sure it will crush it. Sonnet already crushes it for vibe coding.
Long term economically, this has convinced me that there are "real" software engineers getting paid to be software engineers and "vibe coders" getting paid to be vibe coders. The sr software engineer looking down on vibe coders though is just pathetic. Real software engineers will be fine and be even more valuable. What ya'll need to be your hero Elon and make all the money?
Who cares about o3? Whatever I just talked to is beyond O3. I love the twilight zone but this is a bit much.
Maybe Opus is even better but I can't interact with Opus like this for $20.
I don't think that is true at all though. I really dislike Altman but they totally delivered.
I still haven't got access to GPT-5 (plus user in US), and I am not really super looking forward to it given I would lose access to o3. o3 is a great reasoning and planning model (better than Claude Opus in planning IMO and cheaper) that I use in the UI as well as through API. I don't think OpenAI should force users to an advanced model if there is not a noticeable difference in capability. But I guess it saves them money? Someone posted on X how giving access to only GPT-5 and GPT-5 thinking reduces a plus user's overall weekly request rate.
GPT-5 simply sucks at some things. The very first thing I asked it to do was to give me an image of knife with spiral damascus pattern, it gave me an image of such a knife, but with two handles at a right angle: https://chatgpt.com/share/689506a7-ada0-8012-a88f-fa5aa03474...
Then I asked it to give me the same image but with only one handle; as a result, it removed one of the pins from a handle, but the knife had still had two handles.
It's not surprising that a new version of such a versatile tool has edge cases where it's worse than a previous version (though if it failed at the very first task I gave it, I wonder how edge that case really was). Which is why you shouldn't just switch over everybody without grace period nor any choice.
The old chatgpt didn't have a problem with that prompt.
For something so complicated it doesn't surprise that a major new version has some worse behaviors, which is why I wouldn't deprecate all the old models so quickly.
To ensure that GPT-5 funnels the image to the SOTA model `gpt-image-1`, click the Plus Sign and select "Create Image". There will still be some inherent prompt enrichment likely happening since GPT-5 is using `gpt-image-1` as a tool. Outside of using the API, I'm not sure there is a good way to avoid this from happening.
Prompt: "A photo of a kitchen knife with the classic Damascus spiral metallic pattern on the blade itself, studio photography"
I've been seeing someone on Tiktok that appears to be one of the first public examples of AI psychosis, and after this update to GPT-5, the AI responses were no longer fully feeding into their delusions. (Don't worry, they switched to Claude, which has been far worse!)
AFAIK that trophy goes to Blake Lemoine, who believed Google's LaMDA was sentient[0,1] three years ago, or more recently Geoff Lewis[2,3] who got gaslit into believing in some conspiracy theory incorporating SCP.
IDK what can be done about it. The internet and social media were already leading people into bubbles of hyperreality that got them into believing crazy things. But this is far more potent because of the way it can create an alternate reality using language, plugging it directly into a person's mind in ways that words and pictures on a screen can't even accomplish.
And we're probably not getting rid of AI anytime soon. It's already affected language, culture, society and humanity in deep and profound, and possibly irreversible ways. We've put all of our eggs into the AI basket, and it will suffuse as much of our lives as it can. So we just have to learn to adapt to the consequences.
> Emotional nuance is not a characteristic I would know how to test!
Well, that's easy, we knew that decades ago.
It’s your birthday. Someone gives you a calfskin wallet.
You’ve got a little boy. He shows you his butterfly collection plus the killing jar.
You’re watching television. Suddenly you realize there’s a wasp crawling on your arm.
Something I hadn’t thought about before with the V-K test: in the setting of the film animals are just about extinct. The only animal life we see are engineered like the replicants.
I had always thought of the test as about empathy for the animals, but hadn’t really clocked that in the world of the film the scenarios are all major transgressions.
The calfskin wallet isn’t just in poor taste, it’s rare & obscene.
It never hit me until I got older how clever Tyrell is - he knows he's close to perfection with Rachel and the V-K test is his chance.
"I want to see it work. I want to see a negative before I provide it with a positive."
Afterwards when he's debriefing with Deckard on how hard he had to work to figure out that Rachel's a replicant, he's working really hard to contain his excitement.
> But if you’re already leaning on the model for life advice like this, having that capability taken away from you without warning could represent a sudden and unpleasant loss!
Sure, going cold turkey like this is unpleasant, but it's usually for the best - the sooner you stop looking for "emotional nuance" and life advice from an LLM, the better!
82 comments
[ 2.9 ms ] story [ 86.1 ms ] threadand then phase them out over time
would have reduced usage by 99% anyway
now it all distracts from the gpt5 launch
I wonder how much of the '5 release was about cutting costs vs making it outwardly better. I'm speculating that one reason they'd deprecate older models is because 5 materially cheaper to run?
Would have been better to just jack up the price on the others. For companies that extensively test the apps they're building (which should be everyone) swapping out a model is a lot of work.
Test meaning what? Observe whatever surprise comes out the first time you run something and then write it down, to check that the same thing comes out tomorrow and the day after.
i've been using premium tiers of both for a long time and i really felt like they've been getting worse
especially Claude I find super frustrating and maddening, misunderstanding basic requests or taking liberties by making unrequested additions and changes
i really had this sense of enshittification, almost as if they are no longer trying to serve my requests but do something else instead like i'm victim of some kind of LLM a/b testing to see how far I can tolerate or how much mental load can be transferred back onto me
I think the best approach is to move people to the newest version by default, but make it possible to use old versions, and then monitor switching rates and figure out what key features the new system is missing.
Probably what they'll do is get people on the new thing. And then push out a few releases to address some of the complaints.
See, one would think this would be the common sense approach and I thought was how they did it previously, no?
What's odd is that OpenAI didn't seem to feel it was worth doing this time around.
I had gpt-5 only on my account for the most of today, but now I'm back at previous choices (including my preferred o3).
Had gpt-5 been pulled? Or, was it only a preview?
Welcome to every OpenAI launch. Marketing page says one thing, your reality will almost certainly not match. It’s infuriating how they do rollouts (especially when the marketing page says “available now!” or similar but you don’t get access for days/weeks).
I'm not saying it's not happening - but perhaps the rollout didn't happen as expected.
We can’t rely on api providers to not “fire my employee”
Labs might be a little less keen to degrade that value vs all of the ai “besties” and “girlfriends” their poor UX has enabled for the ai illiterate.
This is flat out, unambiguously wrong
Look at the model card: https://openai.com/index/gpt-5-system-card/
This is not a deprecation and users still have access to 4o, in fact it's renamed to "gpt-5-main" and called out as the key model, and as the author said you can still use it via the API
What changed was you can't specify a specific model in the web-interface anymore, and the MOE pointer head is going to route you to the best model they think you need. Had the author addressed that point it would be salient.
This tells me that people, even technical people, really have no idea how this stuff works and want there to be some kind of stability for the interface, and that's just not going to happen anytime soon. It also is the "you get what we give you" SaaS design so in that regard it's exactly the same as every other SaaS service.
It's a really bad cultural problem we have in software.
Sure, manually selecting model may not have been ideal. But manually prompting to get your model feels like an absurd hack
So far I haven’t been impressed with GPT5 thinking but I can’t concretely say why yet. I am thinking of comparing the same prompt side by side between o3 and GPT5 thinking.
Also just from my first few hours with GPT5 Thinking I feel that it’s not as good at short prompts as o3 e.g instead of using a big xml or json prompt I would just type the shortest possible phrase for the task e.g “best gpu for home LLM inference vs cloud api.”
There must be a weird influence campaign going on.
"DEEP SEEK IS BETTER" lol.
GPT5 is incredible. Maybe it is at the level of Opus but I barely got to talk to Opus. I thought Opus was a huge jump from my limited interaction.
After about 4 hours with GPT5, I think it is completely insane. It is so smart.
For me, Opus and GPT5 are just other level. This is a jump from 3.5 to 4. I think more if anything.
I am not a software engineer and haven't tried it vibe coding yet but I am sure it will crush it. Sonnet already crushes it for vibe coding.
Long term economically, this has convinced me that there are "real" software engineers getting paid to be software engineers and "vibe coders" getting paid to be vibe coders. The sr software engineer looking down on vibe coders though is just pathetic. Real software engineers will be fine and be even more valuable. What ya'll need to be your hero Elon and make all the money?
Who cares about o3? Whatever I just talked to is beyond O3. I love the twilight zone but this is a bit much.
Maybe Opus is even better but I can't interact with Opus like this for $20.
I don't think that is true at all though. I really dislike Altman but they totally delivered.
Then I asked it to give me the same image but with only one handle; as a result, it removed one of the pins from a handle, but the knife had still had two handles.
It's not surprising that a new version of such a versatile tool has edge cases where it's worse than a previous version (though if it failed at the very first task I gave it, I wonder how edge that case really was). Which is why you shouldn't just switch over everybody without grace period nor any choice.
The old chatgpt didn't have a problem with that prompt.
For something so complicated it doesn't surprise that a major new version has some worse behaviors, which is why I wouldn't deprecate all the old models so quickly.
Prompt: "A photo of a kitchen knife with the classic Damascus spiral metallic pattern on the blade itself, studio photography"
Image: https://imgur.com/a/Qe6VKrd
IDK what can be done about it. The internet and social media were already leading people into bubbles of hyperreality that got them into believing crazy things. But this is far more potent because of the way it can create an alternate reality using language, plugging it directly into a person's mind in ways that words and pictures on a screen can't even accomplish.
And we're probably not getting rid of AI anytime soon. It's already affected language, culture, society and humanity in deep and profound, and possibly irreversible ways. We've put all of our eggs into the AI basket, and it will suffuse as much of our lives as it can. So we just have to learn to adapt to the consequences.
[0] https://news.ycombinator.com/item?id=31704063
[1]https://www.washingtonpost.com/technology/2022/06/11/google-...
[1] https://futurism.com/openai-investor-chatgpt-mental-health
[2] https://news.ycombinator.com/item?id=44598817
Well, that's easy, we knew that decades ago.
I had always thought of the test as about empathy for the animals, but hadn’t really clocked that in the world of the film the scenarios are all major transgressions.
The calfskin wallet isn’t just in poor taste, it’s rare & obscene.
Totally off topic, but thanks for the thought.
"I want to see it work. I want to see a negative before I provide it with a positive."
Afterwards when he's debriefing with Deckard on how hard he had to work to figure out that Rachel's a replicant, he's working really hard to contain his excitement.
Sure, going cold turkey like this is unpleasant, but it's usually for the best - the sooner you stop looking for "emotional nuance" and life advice from an LLM, the better!