The companies aren’t changing anything. LLM outputs are just more random than people realize. Run the same prompt 10 times if you really want to know how well they can answer.