The only metric that GPT 4.5 seems slightly better on is hallucination avoidance, and that’s only compared to other GPT models. It’s notably worse at software and math problems, it struggles with continuity, and is just on par with cheaper models for research. Given the GPU time and budget to get to this milestone, this seems like a *massive* flop from ChatGPT. DeepSeeks newest model is due in March. Claude delivered the 3.7 upgrade in half the time and knocked it out of the park on software, logic and inference. We’re really seeing the idea of throwing more computing power at a model reach entirely diminished returns.