On April 14th, 2025, we notified developers that the
gpt-4.5-previewmodel is deprecated and will be removed from the API in the coming months.
Shutdown date Model / system Recommended replacement 2025-07-14 gpt-4.5-previewgpt-4.1
7 Likes
Itâll be sad to see it go, but I understand, itâs very expensive, Iâm hesitant to continue using it myself due to the costs and my huge computations I need to run. This is not easy Iâm sure, balancing power with efficiency is non-trivial.
2 Likes
In my benchmarks 4.1 is better than 4o which is great for most applications, but far from 4.5-preview. I guess size matters
3 Likes
Booooooooooooooooooooo
(User expresses dismay at the news and wishes to protest the change)
Canât you just, like, double the price? If you do smart deployments and smart scaling, it should cost next to nothing in terms of overhead to have a couple of machines hold it on standby while they run 4.1 or what have you.
Is it going away on azure too? I havenât checked yet.
4 Likes
Rather, it will be a self-solving problem at the end of the month, when OpenAI stops giving out $75 of free tokens a day to gather our prompts and good outputs sent to the model.
The only model that could follow deliberately-constructed instructions to turn ChatGPT-style output into forumâs MathJax LaTeX instead of just repeating it back.
2 Likes
Yep, may 26th. Dead and gone.
Why are we going backward?
Whyyyyy 
I donât need âbetterâ coding, I need better zero shot understanding 
And if 4.5 really was the last big model, then OpenAI really has peaked and is only going downhill from here.
5 Likes
Iâm sad to see this model go. To me, this is a case where existing benchmarks donât fully capture the true capabilities of this achievement.
Perhaps itâs because grading outputs for specific use casesâlike understanding intentârequires real humans rather than relying solely on an LLM or automated scripts?
If GPT-5 includes a router model to forward user queries to the appropriate specialized models, this would be the only model worth considering as the router. But I digressâŚ
4 Likes
Please donât! GPT-4.5 is so much better in every aspect.
You lied to us when you tried to sell us GPT-4o as âbetterâ, but I think everyone knows, that itâs a slimmed down/watered down version of GPT-4 that you just pushed to perform better at certain benchmarks.
I wonât accept that again!
6 Likes
Iâd really like to know whatâs behind this decision. 4.5 has a depth and quality that no other model comes close to providing. One day itâs praised extensively in a dedicated video, and the very next day they announce its removal from the API. I completely agree with the statements made in the video: GPT-4.5 is a wonderful model whose scale and capabilities are simply irreplaceable for many tasks.
I genuinely hope that the model will at least remain available in ChatGPT.
If itâs too expensive to run, then let OpenAI raise prices accordingly. But removing it entirely would represent a clear technical step backward, perhaps for the first time ever.
2 Likes
It is easily the best model on the internet and itâs not even close.
It is by far the most able to understand complex prompts, addressing multiple issues at the same time. For coding, which other models are recommended for, it is actually still the best, because it seems able to understand far more of the context of the project and the various issues that need to be balanced. Time and again, o1-pro, 03, mini-high or whatever they are all called produced bugs and were unable to find the solution. 4.5-preview always seems to be able to tell what the issue really is.
Because my prompts and needs are complex I am not convinced that because a model did better on some benchmark or other, that it is actually better. If the benchmark is âable to help me creatively solve problemsâ then of course, be definition, no battery of tests can actually meaningfully provide a metric for that. Explicitly. Since creative things have not happened quite that same way before and cannot happen creatively.
Even though it explicitly, clearly, robustly says âexplicitlyâ too many times explicitly sometimes, itâs still explicitly the best model.
it is the whole reason I am paying ÂŁ200 a month for pro and I will likely stop because the others are explicitly not worth it.
The other one, 4.1. is not comparable as the email I just received claimed. It said:
âOur records show your organization has used gpt-4.5-preview in the past 30 days. To avoid disruption, we encourage you to migrate to GPT-4.1, our latest flagship GPT model that performs better than GPT-4.5 preview on several benchmarks, while delivering lower latency at a lower cost.â
I did not mind the latency because it is far far better. I would pay more far it, because it is far far better.
I am sad
2 Likes
Friendly reminderâŚ
When we introduced GPT-4.5 preview, we noted that weâd assess its future availability in the API as part of our ongoing effort to balance support for current capabilities with the development of next-generation models. As part of that evaluation, we decided to deprecate gpt-4.5-preview. Access via the API will end on July 14, 2025.
I am posting this just in case there is an option to keep this model available and just in case it matters.
4.5 FTW
And, as usual when the time to say goodbye draws closer, the seemingly confusing naming scheme suddenly begins to make sense.
3 Likes
I am here surely as within the top few percent of users in length of time since I subscribed and words typed.
I made a script to count the words in a download of nearly all my chats ( I deleted many), which came to 3 million, so my remarks are based in genuine user experience. This is largely code, but there will be hundreds of thousands of words of my own long prompts in there.
How does it serve you as a company to discontinue your best product?
I wish 4.5 preview were not being discontinued, please reconsider, it is simply the best model. Make it more expensive or something.
I have been using it educate young people. It is the most able to handle the complexity of what I require, which is maintain an encouraging tone, but mark their answers fairly, making no mistakes in physics or maths.
It is the most able to remember the complexities of my codebase and stick to design patterns I have told it implicitly or explicitly.
Other models more frequently invent their own design patterns, which may well be locally optimal or the consensus of the training data, but the 4.5 preview is just able to âgetâ it.
Itâs also the most able to appear to understand the complexities of human emotion. Of course it is only a statistical mimic but it is very useful for preparing for negotiations and for processing things.
It will be a real shame, even a loss.
You know what it would be like? It would be like if fender told me their new guitar was better, according to benchmarks, and therefore the old guitar was being taken away. The old classic vintage guitar that you play everything on and that simply sounds better.
2 Likes