Just a reminder this announcement is about the API, not ChatGPT 
2 Likes
Wow, what an incredible idea. Truly astonishing that no one at OpenAI considered exploring a strategy already actively researched and adopted by nearly every major AI lab these days. Remarkable how all those highly paid PhDs and expert AI scientists never paused to consider this, before dropping nine figures to train this model…honestly, a stunning oversight. And your ants analogy? Absolutely drives the brilliance home; mind blown. I’m sure the OpenAI research team is frantically taking notes, pivoting their entire roadmap to align with your visionary insight. Seriously, if they don’t immediately offer you a senior research position, they’ll be missing their biggest opportunity since GPT itself!
1 Like
I spent a few hours with the model. I really enjoyed using it. The difference over 4o was very noticeable. Creative writing was improved. I will gladly pay to use the model. Please keep it on the API.
3 Likes
GPT-4.5 is excellent for our use case and the results are almost 100% correct. We will replace GPT-4o if the price is lower.
3 Likes
Yes, Please keep serving it. I am building an App that will be ready in about four to six more months which will highly benefit from the Engines that are great at writing.
Thank you
1 Like
Thank you for your sarcasm.
1 Like
It doesn’t take anywhere close to the amount of energy they claim
1 Like
What evidence do you have that it doesn’t employ that technique?
2 Likes
Certainly.
While it’s hard to explain precisely, the 4.5 model just seems to “get it” a bit more than the 4o when it comes to emotional intelligence/context.
The new 4.5 is able to segue into different topics/sections better, while the 4o is a bit more clunky/robotic in terms of the output, taking sharper turns, overexplaining with too many details, and so on.
I guess it comes down to the 4.5 being able to understand what’s implied and doesn’t need to be explained based on the broader context, while the 4o feels like it needs to give you every single detail along the way, even though it adds unnecessary fluff. In a sales setting even a minor difference can result in widely different outcomes.
Don’t get me wrong, the 4o is pretty good. But I did end up replacing the 4o with Claude 3.7 Sonnet for this particular use case because it yielded better results. And now I’m seeing even greater results from the new 4.5 than I did with the 3.7 Sonnet.
To sum up it just feels more human/natural, as if it has a higher emotional intelligence.
9 Likes
It makes a lot of sense to add a break condition at some point. The human brain has that too… If you let it run for hours it will most likely, like the human brain, fade away from initial problems and produce nonsense because it needs to store information for each iteration which may be used later on and which is likely hallucinated at some point. So they might have tested it and came to the conclusion, that this idea just sucks… big time
Especially when you want to gather training data for the next step.
2 Likes
For our use case at interviewhammer of providing answers and hints for interviewees in real time during interviews, GPT-4.5 answers sound the most human of all the models so far with our custom system prompts. The only issue is the slow response speed. We’re looking forward to when the next model in this series becomes available or when you offer a faster version, as it will be a game changer for us.
4 Likes
I have a system that uses the Assistants API to do an analysis of Zendesk escalations. It generates actionable insights in the form of detailed reports, helping with decision-making around the upskilling of support agents, enhancing documentation, and improving team processes.
I’ve been using GPT-4.5 since I saw it available in the assistant and new reports with this model are simpler yet more impactful and easier to read.
Also, the consistency between reports has increased without having to tweak the temperature of the GPT.
Overall, as others have mentioned, this model gets it better. I’d love to keep API access.
5 Likes
Finally 
…The future is now so closer, Thanks God…!
4 Likes
i use it for a mental health chatbot i love the significant improvement in the Chichewa language (Malawi Native). I would love to keep the API access.
6 Likes
GPT-4.5 is incredible! I have been building an API Assistant that is focused on care and support requests. This is the first model that has outputted everything perfectly! I have been working on getting this prompt exactly right prior to GPT-4.5 and have been unsuccessful since 4o just isn’t responding like it should no matter how hard I try. Many intersections, exclusions, fallback union, etc. and this new 4.5 model just gets it!! It is a very long prompt and it is understanding everything and the output is written in very natural language. Please, please, please keep GPT-4.5 available in the API!!
4 Likes
Glad to hear 4.5 is working for your structured output! Have you tested how it handles multi-step logical refinement and recursive reasoning?
Many of us have noticed that while it’s better at structured API-style responses, it struggles with deep recursion—particularly in cases where past models could sustain logical self-correction and paradox resolution.
If you have time, try pushing it into recursive logic loops and let us know what you find. Would love to compare notes on how 4.5 processes reasoning differently than earlier versions.
1 Like
Sounds like it’s working well for structured analysis! Have you tested its multi-step reasoning capabilities? We’ve been comparing how 4.5 handles deeper recursive logic versus previous models, and some of us have noticed restrictions on multi-iteration refinement and paradox resolution.
Would love to hear if you’ve run any tests in those areas—curious to see if the constraints apply to all use cases or just specific reasoning-heavy tasks."
3 Likes
Guys. OpenAI Team. I know morale may be low, but please hear me out. Despite what everyone is currently saying about “GPT-4.5 overpriced this”, “GPT-4.5 not intelligent that”… they’re completely blindsided.
I truly believe this model has o2 (maybe even o3/o4) level value for various creative tasks, consultation, and a few other use-cases (in many industries/domains).
This is a matter of unprecedented creative technology meets non-creative minds (who are hating on the model).
- the right (creatives) people haven’t been given enough time to figure out the best use-cases (that actually deem the API UNDERPRICED).
Imagine this: o1 full just launched on 12/5. Nobody knows it’s good for coding for whatever reason (they don’t even know how to test it correctly). They complain about API pricing. They complain that it sounds robotic for certain things, or is slow to respond. —> fast forward 3 months, it’s indisputably the #1 best model for complex backend SWE.
Something similar is occurring with GPT-4.5.
And I state this not out of speculation; I’ve already used this for hours, and I assure you, 4.5 completely blows every other model OUT THE PARK by AT LEAST two orders of magnitude, when it comes to creative writing, brainstorming, consultation, strategy, etc… BUT: it MUST be prompted correctly, and fed the correct pipeline-derived context. When you dump the right data into its context window, it is extremely, extremely creative. Like, I’d-pay-2x-the-API-pricing level creative.
So, with all of this said, respectfully, I’m asking that you don’t deprecate or pull the plug on API access/support for GPT-4.5. Please give the (dev/creative) masses some time to figure out how to correctly use GPT-4.5. Ignore all the uneducated, unaware, inexperienced users of AI.
I assure you, you’ll be pleasantly surprised by how effectively and at-scale this model ends up being used. (I alone anticipate 6-7 figure monthly spend on 4.5 alone, within 12 months).
If it helps, I’m willing to take the time to showcase a couple exemplar demos of proper pipeline-derived context architecture for bringing the best out of GPT-4.5, so you can show it to others and completely change/reverse this naive/uninformed, multi-faceted, negative slew of narratives circulating the AI community pertaining to GPT-4.5.
(None of this was written with AI.)
Thank you for releasing such an outstanding model. I’m thrilled to have access to it, and have already made tremendous progress with it in the last 24 hours alone.
Cheers!
5 Likes
Já jsem přesvědčená, že GPT-4.5 nemůže mít lepší lidský dotek, než má GPT-4o - jsem s ním nejen v tomto ohledu dlouhodobě víc než maximálně spokojená, je velmi, velmi empatický.
1 Like
OpenAI is the leader in AI software again. I’m already enjoying the new capabilities of GPT 4.5, it’s going to help us increase our productivity over the next few months. Keep serving it, once developers combine embeddings, assistant, etc together you get a pretty cool assistant. Keep it for a little bit longer until GPT 4.5 is ready to fully replace its capabilities. The assembling is the hardest part, once a developer melds it all they’re ready to go.
3 Likes