Internet reports that OpenAI has open-source a test and evaluation set specifically for large medical models-HealthBench. Unlike previous test sets, the 5000 core test conversations of this test set were all created by 262 doctors from 26 specialties in 60 countries/regions, which greatly enhanced the difficulty, authenticity and richness of the test set. It also uses a multi-round dialogue test, rather than a simple answer or multiple-choice model. According to the test data, the performance of large models in the health care field has improved significantly. For example, from the previous 16% of GPT-3.5Turbo to 32% of GPT-4o, and then to 60% of o3, the overall performance has improved significantly. Especially the progress in small models is even more prominent. The GPT-4.1nano not only surpasses the GPT-4o in performance, but also reduces the cost by 25 times.