I’m considering the OpenAI Batch API for a job where a few bad inputs shouldn’t invalidate thousands of good ones. If individual requests have malformed parameters or hit a model-side error, does the batch still finish and return results for the successful requests?

I’m also unsure how best to match each result back to its input and distinguish a request failure from a batch-level failure. If you’ve handled mixed-success batches in production, what did the output look like?