In our live production system today we have seen a case where the model (gpt-5-mini-2025-08-07) obviously generated a response based on the previous question in the conversation thread, rather than to the question that was asked. We are using previous_response_id to thread the conversation. In other words, we are not adding history into the input explicitly – only implicitly by referring to a previous response. We are using the file-search API. Looking at the full response data for the response and for its predecessor, there’s clearly something wrong. The response clearly answering the previous question (not quite verbatim to the previous response but very similar), while the current question was completely different. In addition, there are two annotations that refer to files that do not even appear among the search results from the file-search tool, but do refer to those in the previous response.
I have the raw response objects if they are needed for troubleshooting.
Has anyone else been experiencing this problem? It’s not easily repeatable, but it has happened multiple times today that we’ve noticed.
As you can imagine, this is a serious threat to our perceived quality on our public-facing service. We’d appreciate help on finding a solution promptly.