OpenBioRQ reveals that agent-based AI models fail on approximately 40% of complex biomedical research questions and paradoxically stop using their tools on difficult tasks, despite these tools being most critical.