On the Linux.do forum, a user conducted a web search capability test on mainstream AI models Claude, GPT, and Gemini, evaluating hallucination rates for questions with scarce information sources. The results showed that Claude Sonnet 4.5 performed best with a 0% hallucination rate, obtaining correct information in just three search rounds; GPT 5.2 had a 70% hallucination rate with low search efficiency; Gemini 3 Pro had a hallucination rate exceeding 90% with poor search results. The author emphasized that Claude is far ahead in tool usage capabilities, such as project management and file operations, and has switched from GPT to Claude as their primary tool. The article calls on AI companies to strengthen tool integration, enhance productivity, and break through model bottlenecks. This test provides practical reference for AI users, revealing performance differences and future development directions among models.
Original Link:Linux.do













