I’m considering the OpenAI Batch API for a large set of independent classification jobs. The lower cost is appealing, but I need to know what happens when only some records fail—for example, because one input is malformed or a request hits a model-specific limit.

Does the completed batch make it straightforward to match each failed result back to its original request and retry just those records? I’d especially like to hear how people handle individual request failures without rerunning successful work.