Gemini 3 PRO Logic: A BEAST!

Key Takeaways

Hello community. Gemini 3 Pro is here. So, let's test it out here with my classical test here. Let's see how it behaves. We agree to everything. Beautiful. So, let's have a look. Generating. Okay, it's a sinking model, which is great. As you see, we just have to wait here a little bit. So, now I'm really interested. How good is Gemini 3 Pro compared to Gemini 2.5 Pro? What is the jump that we can expect here in causal reasoning, in pure logical argumentation? I'm not talking about how to set up a creative project. No. This is mathematics, physics, chemistry, science. How good are you? Can you really do a complex logical sequence and solve it? Gemini 3 Pro is available. Unbelievable. So, let's have a look. Here is the optimal solution to reach floor 50. Really? The exact sequence. Okay. Hidden reason, the shortest sequence, eight button presses. This is absolutely top result. This is the best result we achieved. Eight button presses followed by emergency exit. A B C B in one emergency exit. Step-by-step state table. Beautifully. Look, energy packets, token flags, code cards. Everything is presented. Nice. Exit emergency exit at floor 29. Green tokens. Yes. Final score, the total presses are eight plus the emergency exit. The final floor is 50. The goal is reached. Final resources, we have enough energy packages, we have all the tokens. Yes, the code cards are collected. Zero random traps. Beautiful. We avoided here all the trap. Beautiful. Proof of optimality to reach floor 50 in less than 20 moves. Utilize emergency exit at floor 29 is a dominant strategy. Correct. Climbing manual to 15. Yeah, that's not the way to go. Code acquisition. Green card, red card. Optimal order. My goodness, this is This is outstanding. Step minimization from 0 to 15. So, it's subdivided this and then you see the step from 15 to 29. Real nice. We must execute this. Analysis for conclusion, minimum viable moves, 4 + 4 = 8. No shorter sequence exists. Done. Wow, impressive. Absolutely impressive. Okay, you know what we're going to say. I'm going to say verify result and maybe you'll find a shorter path. Uh, what do you think about it? No? Maybe? Gemini 3 Pro? Uh, wow, what a model. I mean, brilliant. Absolutely loving it. I have no idea how it achieved this, if it was coding in the background, yes or not, but the result, I can tell you this is the correct result. Because, yes, it outperforms all the other models we had until now. It outperforms GPT-5 and it outperforms here everything else that I tested. Wow. It looks good. So, let's have a look at Gemini 3 Pro here. Verify your result. Maybe find a even shorter path. The previous solution of eight button presses is correct. Optimal extensive search confirms that there is a seven button press exist. It fails the resource constraint due to penalties and the lack of refunds. Here's the verification of the final optimal solution. Token bottleneck. Okay, I see. Why seven moves fail? Seven moves forces two consecutive A presses. We have a penalty. Yes, minus one token. Yeah, okay, the token don't add up. Eight moves succeeds. Eight move path inserted an extra step to land at floor 11. Grants one token. No penalty. Result success. Optimal solution. Again, eight button presses. With the emergency exit. Final stats. Total presses, eight. Energy enough, token enough, code green card, everything that we need. Trap zero. Proof of Pareto optimality exists because solution is optimal. Okay. Now, I don't know how to improve this test because this is really just simply amazing. And by the way, you know, I just tested days ago the new GPT-5.1 here and this is the video here. And as you can see on the left side, we had GPT-5.1 and GPT-5.1 told me this is impossible to construct a legal run reaching here floor 15 in less than 20 button presses. Therefore, the current rule and the strict floor bound this puzzle is infeasible as stated. So, only if I'm willing to relax one constraint, then GPT-5.1 can solve it. And GPT-5 chat, the current chat just days ago told me, "Hey, you know what? I found a legally bound optimal elevator run that reaches the floor 15, 18 to 20 presses." So, even the chat GPT-5 chat found a solution at the limit of what is allowed, 18 to 20 presses. So, just that you can compare another performance of the classical non-sinking GPT-5.1 with the new Gemini 3 Pro. And I don't know if the the Gemini 3 Pro was even executed here in the high model version. This is it for today. I hope you are giving you here a very first impression. Would be great if you subscribe and I shall see you in my next video.