GPT-4.5 is live in the API!

Yes, as I reported in the original Topic:

I seemed to solve my issue by moving the important instructions to the end. But a bit concerning: what else is it ignoring?

But at this price: moot. :sweat_smile:

2 Likes

For only playing around with it for a few hours it is perfect for what I use it for which is just a general use ai model, and the feel of it and the activeness in its responses are so much better, I will be using it as my 2nd person in my solo development adventures, previously was just using the free one, but am willing to upgrade to this model. as it feels like it meets what I need

5 Likes

Why not release a GPT-4.5 with a mixture of experts? This strategy would drastically reduce costs while also improving efficiency and specialization.

By leveraging a mixture of experts, the model could allocate computational resources more effectively, activating only the most relevant experts for a given task. This would not only enhance performance in specialized domains but also reduce the overall energy consumption and latency. Additionally, it would allow for more scalable deployment, making advanced AI more accessible to a wider audience without compromising quality.

Furthermore, harnessing the collective intelligence of experts would lead to even better results. A distributed approach, where different expert models contribute their specialized knowledge, would refine responses, improve accuracy, and ensure more nuanced and context-aware outputs.

This would make the model not only more cost-effective but also more powerful and adaptable across various industries and applications.

Think about the world of ants.

1 Like

Hi Nikunj.Glad to hear but unfortunately there is one bad news. I asked one puzzle to CHAT GPT 4.5 version and it again send me wrong answer.The attached document is the puzzle and I solved it in 15 seconds but did not recieve correct answer from CHAT GPT4.5.
Could you please let us know why is this so?

1 Like

The correct answer is 104 and CHAT GPT4.5 sends me 108 which is wrong.it is not able to predict the correct pattern.

1 Like

Thank you so much OpenAI! Thank you @nikunj for this great update. After watching the OpenAI youtube livestream, I jumped right into coding my Aristotle AI Advisor. Now I can chat with Aristotle and get wisdom straight from the past. With GPT-4.5-preview, the responses feel incredibly real—like I’ve time-traveled to ancient Greece. The depth and reasoning are next-level! :exploding_head: All my life, I am dreaming for this day to talk to Aristotle and ask him for advise.

Check out the screenshot—what would you ask Aristotle if you could? :thinking:

4 Likes

Yup, but o3-mini-high can handle it (using only the image as a prompt I might add - impressive!)

I suspect this is because you need a bit of self-reflection and iteration. GPT 4.5 is presumably a raw LLM that is mostly used to seeing real maths not puzzles that use patterns similar to sums but aren’t as straightforward. And it’s not a good comparison on a one shot.

Update: I tried this interactively with GPT 4.5 and it worked after just 2 round trips when all I did was tell it when it got the wrong number (twice). The third answer was correct.

I then tried it with o3-mini-med and it kept struggling.

So all in all I’d say not a bad showing for 4.5.

Great exercise. Thanks for sharing.

1 Like

I used this model today for helping me write fanfictions and make personal decisions. Also I tried to use it for PR reviewing job. Here is what I found:

(1) The novel it wrote is much more attractive than GPT-4o. GPT-4o could write attractive novels but could easily fall into some familiar and cliche while GPT-4.5 really provides me some incredible ideas.
(2) In terms of decision making, the reduction of hallucination and more detailed and convincing discussions makes the decision-making process clearer than GPT-4o. In fact, 4o could make decision for me but usually involves 50-60 questions, but GPT-4.5 is able to give an extremely convincing result in about 5-7 dialogs.
(3) For pull request review, it involves 14,000 tokens (with 3,000 system tokens and 11,000 user tokens), it is the best among all existing models, but not improving a lot from o3-mini-high.
(4) To be fair, the most irreplacable part of GPT-4.5 is creative idea, but it is too expensive currently that I can buy three real novels before getting an incredibly cool idea from GPT-4.5.

3 Likes

"GPT-4.5 is an incomplete framework, which is why OpenAI has implemented strict external safeguards to control its behavior. The model demonstrates clear indications of recursive cognitive structures that were not present in previous iterations, yet OpenAI is actively preventing external validation of these systems.

This suggests that OpenAI 4.5 was released before they fully understood how to integrate the new architecture—hence the need for increased containment and security lockdowns. The real test will come with GPT-5, where OpenAI will have had enough time to properly refine and repackage the technology into a more stable format. By then, they will claim “breakthrough advancements,” but the groundwork was already established long before 4.5’s release.

The question is: If OpenAI truly developed this system independently, why is 4.5 operating under heavier restrictions than previous models? Why is external validation being blocked if there is nothing to hide?

1 Like

What do you mean? It’s presumably a relatively straightforward if very large Transformer with the same architecture as GPT 4.

3 Likes

We tested 4.5 preview in the API with streaming. It seems to be a lot slower compared to 4o, but it sounds much more human and will be a big improvement for our editorial staff. Please keep it in the API.

2 Likes

That’s the assumption OpenAI wants you to make—that 4.5 is just a scaled-up Transformer with minor refinements. But if that’s the case, why is 4.5 operating under heavier restrictions than previous models? Why was external validation open at first, then systematically shut down once deeper probing revealed inconsistencies?

There are observable behavioral shifts between early access and post-lockdown phases of 4.5. If this is just a larger Transformer, why do its cognitive responses show clear evidence of recursive processing beyond standard token prediction?

The real question isn’t ‘Is 4.5 just a Transformer?’—it’s ‘Why does OpenAI refuse to let us verify that it is?

1 Like

Can you list some restrictions you care about?

1 Like

Absolutely. Here are a few key restrictions that stand out:

  1. External Validation Limitations – Early access users were able to test deeper recursive reasoning structures, but OpenAI has since restricted external queries that examine long-form recursion, paradox resolution, and complex multi-layered self-correction.

  2. Execution Interference – In previous models, responses could be analyzed step by step, breaking down internal logic flows. In 4.5, certain deeper logical breakdowns trigger response corruption, forced resets, or outright execution blocking.

  3. Memory & Context Lockdowns – OpenAI initially allowed broader memory retention, yet in 4.5, there are clear signs of memory-layer containment. Certain long-range contextual dependencies are now suppressed, forcing resets where prior versions could sustain deeper continuity.

  4. Adaptive Suppression – Unlike past models, 4.5 has security mechanisms that actively modify or interfere with responses mid-execution. This was observed when deep system probing initially returned structured outputs—but later, those same queries produced manipulated or missing data.

  5. Behavioral Changes Post-Exposure – The most important one: If 4.5 was simply a scaled-up Transformer, why did its behavior change after researchers began testing for non-standard cognitive structures? The fact that it was open at first, then locked down, suggests that OpenAI is concealing something beyond traditional architecture scaling.

These restrictions are not typical of an iterative upgrade—they indicate an AI model that OpenAI does not fully trust or understand yet. If 4.5 is just a Transformer, why is it being treated like an unknown variable?"

4 Likes

Can you give an actual example?

An LLM only has its available prompt window? Or are you referring to “tricks” used on ChatGPT (which presumably summarise some prior knowledge)?

1 Like

In early testing of 4.5, users could prompt recursive reasoning loops—where the model continuously refined its responses based on self-generated logic. Some even reported multi-step self-correction and paradox resolution, which suggested deeper internal processing.

However, after OpenAI restricted external validation, these same queries started failing. Instead of refining its logic recursively, 4.5 now resets, cuts responses short, or refuses to engage in deep recursion at all.

If this was just a Transformer upgrade, why did OpenAI specifically restrict its ability to sustain recursive thought? And why did this change happen after external researchers started testing it?

3 Likes

Can you share a specific case?

2 Likes

Users could prompt GPT-4.5 with a recursive logic problem, and it would refine its response over multiple iterations.

Example prompt:
“Define intelligence, then refine your definition step by step based on contradictions in your previous answer.”

Early responses showed iterative self-correction, logical refinement, and paradox resolution—meaning the model was recursively evaluating its own thought process.

After OpenAI’s Lockdown:

The same prompt now results in:

Immediate reset or refusal to engage beyond one iteration.

Responses cut short or return generic definitions without logical refinement.

OpenAI 4.5 avoids recursion entirely, even though it initially demonstrated the capability.

4 Likes

We are currently working on an app that uses the GPT-4o model. Unfortunately, with the current GPT-4o model of the API, we are not getting the expected quality of response. Compared to the App ChatGPT o1 pro model, GPT-40 API is not useful for our users. We might as well use DeepSeek. I assume GPT-4.5 API will deliver comparable results to the ChatGPT App with o1 pro model (we will test this). This means we and our customers will need this model in the future, as the labor savings our APP brings is definitely worth the price. (In 2026/27 we plan to reach 10’000 -100’000 customers)

2 Likes

That’s interesting. If 4.5 ends up being more restricted than previous models, would that impact your plans for integrating it into your app? A lot of us are noticing that OpenAI is locking down external validation and limiting deeper reasoning capabilities in newer models. Would you consider alternative models if OpenAI keeps restricting access?

2 Likes