This is why I migrated my apps that need an LLM to Gemini. No model degradation so far all through the v2.5 model generation. What is Anthropic doing? Swapping for a quantized version of the model?