GPT3.5 couldn't solve correctly all my sophomore level programming class for mechanical engineer problems
GPT4 could solve every problem from the programming class, but not from my senior-level intro to robotics class
GTP4o can solve every problem from my senior-level intro to robotics class (with some very minor syntax errors needing to be fixed). But, it couldn't solve any theorem/proof problems from my graduate level robot kinematics and dynamics class
GPT o1 can solve every problem from all those classes mentioned above, including giving detailed explanations of the solutions. It still could only do about 50% of the theorem/proof problems correctly, but even on the ones that weren't convincing proofs it gave them the key avenue they needed to follow to do the correct proof.

It is an utter fallacy to claim that these AI models aren't getting better over the last year or two. I think there probably needs to be another innovation above and beyond larger networks and compute power to make the next huge leap, but these are still getting better.