I get what you're saying because this is typically true (this is a strong motivator for my current research) but I don't think it applies here and OpenAI seems to agree with me. Some cases are clear: GPT-5 is clearly better than Llama 3 for example. If there is a sizeable enough difference across virtually all evals, it is typically clear that one LLM is a stronger performer than another.

Experiences aside, Gemini 3 beats GPT-5 on enough evals that it seems fair to say that it is a better model. This appears in line with public consensus, with a few exceptions. Those exceptions seem to be centered around search.