Gemini is objectively exhibiting new behavior with the same prompts and that behavior is unwelcome. It includes hallucinating information and refusing to believe it's wrong.

My question is not whether this is true (it is) but why it's happening.

I am willing to believe the aider community has found that Gemini has maintained approximately equivalent performance on fixed benchmarks. That's reasonable considering they probably use a/b testing on benchmarks to tell them whether training or architectural changes need to be reverted.

But all versions of aider I've tested, including the most recent one, don't handle Gemini correctly so I'm skeptical that they're the state of the art with respect to bench-marking Gemini.