I'm very curious what your prompts are, and whether you're cherry-picking (deliberately or not). I can't reproduce any of your findings with ChatGPT, Gemini, or Gemma 4 (within AI Studio).