"GPT-5.4 HIGH" Cheating? Can it Reason or just Write Code?

Key Takeaways

Hello community. So great to see you. You know, Open EI released just hours ago the new GBD 5.4 and here I test now the high version. We have it here March 5th, 2026. OpenUI introducing GBD 5.4 designed for professional work. Great. I have my own scientific causal reasoning test because I do not trust here those benchmark that are known for three or four years. you know they could dilute here maybe the test the pre-training data set so therefore I created here my own test I have here YouTube playlist and you just go and you select them all and you see exactly the performance exactly this I have done with the GPT 5.4 and you see the result here in this video this was not the high this was the normal GPD 5.4 for that I think a lot of people going to use here just as a default. Now if you go there and I do my test here on areno.ai why because this is a free platform you can go there without paying anything without doing a login or whatever. You choose your model and you have these models selected and you do your testing and without paying anything you get an immediately response by your model or by multiple models and you see which model is best for your task and I want that you can do this you can replicate all the test and you don't have to pay for this now there's a difference in philosophy if you go with openi you see we have a GPD 5.4 a 5.4 for high and X high and so on. If you look at Google, you have a Gemini 3.1. You have only the pro version and this is it with GPD 5.4, you know, they want you that you you search here for the best model and you pay a little bit more maybe. So, no problem. We are going to test now GPD 5.4. Hi, let's go to the live testing. Okay, as you can see we have the same test and we have now GPT 5.4 high. So this is the option that we get now we have a reasoning model which makes much more sense if you go for a reasoning test especially if you want to explore here if GBD 5.4 for high is able to solve here a scientific test. And here we go. I will stay here in real time, but of course for you, you will see now in the next second when the GBD 5.4 high finished here its reasoning process and present us the first solution. Oh yeah. So we have something beautiful. Let's have a look. Okay. So [clears throat] what is it? Button sequence AB ABC A B C FA. This is an excellent solution. Then on floor 29 invoked emergency exit 250. This is beautiful. This is gorgeous. This is Yeah, this is what I was hoping for. So here we have the single steps. Here we have all the mathematical operation that are connected here with the different button presses. Okay. And of seven button presses, we have the emergency exit. We on floor 29 and we have a shortcut to floor 50. Beautiful. So therefore the totals and jacks here we have the toten button presses seven plus the emergency exit. The final resources are from the energy within our constraint. The token are within the constraint. The code held are exactly what we need. The random trap hits is exactly zero. This run satisfies every constraint. It reaches the floor 50. Beautiful. Uses at most 20 presses. Finishes here within the constraint. uses no random trap avoids all forbidden interaction. Beautiful. So we do have here an excellent solution and you know exactly we have to validate this result. So here we go GPD 5.4 high since we have now result and it is just a validation this should be done now much faster. Yes, here's the full validation. Beautiful step-by-step validation. Nice. Okay. Why every step is legal? Okay. I have a detailed explanation. Just stay with you here in real time that we can do this. So the run is valid exactly as stated. Let's have a look if this is true. I'm treating the emergency exit on floor 29 as a special action not as a button press. No problem. Then we have here all the button presses here with all the operation. Beautiful. I have here all the information I'm looking for. This is nice. Exactly. I go to the floor 50. Every step is legal to disord the landings. No illegal happens along the way. Blue code, red code, green code. Yeah. Special collections solution match every goal. Yes. Yes. Yes. Okay. [snorts] Within the limits. Beautiful. No traps. So the run is valid and exactly as stated. This is an excellent result for GBD 5.4. for high. So we have to have here a high reasoning precision for any causal reasoning test. And you know what we do? We now ask, hey, can you find a better solution? So I simply ask, hey, can you find a shorter path of button presses on this elevator test that we go from the floor of zero to the floor of 50? What methods can you apply for this task? And what is the coherent result of yours? So let's see what it finds here on possible ways forward. What or how is the methodology to solve this simple complexity here. Okay, we have a result here. Beautiful. So you remember I asked, can you find a shorter path for button presses? Yes. Under the written rules, I do not find a shorter guaranteed run than seven button presses in the emergency exit. Okay. So the clean way to solve this is the shortest path on the full game state has to include everything. Beautiful. Ah BFS. Okay. Okay. A simple mathematical methodology. Okay. Beautiful. Hand checks as lower bounds because they cut the search sharply. Why seven pushes is optimal? H. First, floor 50 is not reachable any normal button route as the rules are written. Well, what a coincidence when I designed this puzzle. Yeah. So, the only coherent finish is the emergency exit from floor 29. This means the real target is reach floor 29 with red and green. This is correct. This is the absolute essence of my puzzle. Now, for the lower binds, red requires ABC. Green requires this. Yes. So, the different floors. Okay. You can reach 29 exactly two more pushes. F with lockdown on six buses a coherent result. An optimal sequence is AB A B C F A then emergency exit 29 state trace. Okay. Seven button presses plus emergency exit zero traffic. This is the shortest correct answer is still seven button presses on my elevator. This is absolutely the best results that also I achieved with the other models. But now to something interesting. I just got the idea. Hey, wait a minute. If uh GPD 5.5 high tells me, hey, I use the BFS here the bread first search method as the methodology to solve you. Then I just ask, hey, can you apply BFS as a Python code when you map the linguistic version of the task into a Python representation or can you also run the BFS method on a pure linguistic task without any coding [clears throat] since you're an agent? And this is the beauty. GBD 5.4 high is per definition more or less here perfectly equipped to act as an agent. It has access to EI Python environment. And now my question is, hey, did GD 5.4 for high found the solution just by syncing here in linguistic terms or did it just map here to a Python and created the Python code and executed here the mathematical BFS methodology on the Python code and therefore became here this perfect problem solver [snorts] and GBD 5.5 comes back and says here after some time yeah both are possible BFS is not tied to Python it's a way of organizing the search Python makes it explicit and repeatable but in plain language which the same idea can be followed by keeping a queue of candidate explanation and expanding them level by level. Now what you have to know that I built my test exactly to prevent that the system can just find out the right solution in trial and error because I have so many combination I have so many alteration I have so many permutation of possible button presses in my mathematical function that I have when you press a button that a pure trial and error I what it was it was a Gemini 2.5 I think and ran more than 45 minutes and it could not from here find any solution at all. So for a plain language I made my model so complex not to find it. But if of course this model or this agent GBD 5.4 writes a Python code and then just runs the code with a mathematical optimization then of course able to find immediately or within one two minutes here the solution. So here it says I give you an example to map a linguistic causal task into a graph and then it really writes here exactly here a simplified version but you see exactly here graph beautiful everything hypothesis so this is simple this is easy and then we have a BFS causal chain just when you apply this great so what happened now unfortunately GPD 5.4 for high will not tell me exactly what it did. This is the big secret. So either it did a BFS methodology just trying out all the tree combination in the pure linguistic way or it just converted this here to a Python code solved it mathematically and came back with the solution and presented me the solution. I have a feeling since it is such a beautiful agentic system it converted to Python and just executed here a mathematical optimization on the Python code. So maybe this was not a real logical exploration of and this was not just a genius idea how to solve this here as a causal reasoning problem because this was just an agent understanding the complexity is too high for the linguistic solver. Therefore it just right. or maybe it write to any other uh solver out there. The simplest case is Python and just solved it mathematically and GVD 5.4I tells me hey what happened here is simple the language facts were turned into symbols and the course of possibilities were turned into graph node and edges and once the mapping exists bfs works very naturally so therefore I think that GBD 5.4 for high was cheating. I mean not cheating, it is an agent and it has tool calls and whatever. So it has access to mathematical solver to a Python environment of course and it can code this turn it into symbols build a graph and found here a mathematical optimization theorem and solve it mathematically. But this is not a sign that this is a pure linguistic genius because maybe it was just writing code. And you know what other model was just writing code? Guess what? Yes, of course it was here from X. It was here grock 4.1. So you see maybe TPD 5.4 for high found his illusion not by logical deduction but just being an agent and writing code. What are your ideas about this? I would be interested to have your opinion. Anyway, I hope you had a little bit of fun. I hope you enjoyed this video. I provided some new information. It would be great to see you in my next