Also its not clear if the same effect appears on larger models like GPT-5, gemini 2.5-pro and whatever the largest most recent Anthropic model is.

The title is an overgeneralization.