I use the Gemini 2.0 model for translation tasks. When I use my own program to call it through the API, some sentences return errors, and the reasons for the errors are sometimes “prompt is disabled” or “unknown error.” However, when I input these sentences into some large model UI clients (such as Chatbox, Cherry AI, etc.) for invocation, they can be translated normally. I believe that these UI clients do not do too much processing of the API calls. Why do they return results more reliably than when I set the security option to block none?
I have tested this multiple times and it is always the case.