They already have. Claude and OpenAI are not trying to write captcha-defying AI agents. These tests wouldn't hold up as well against proper bot operators who mimic user behaviour. However, the signals are still valid as part of a larger toolset.