I’m considering moving a nightly queue of short text-classification requests to the OpenAI Batch API. The work doesn’t need an immediate response, but I do need a dependable way to match each result to its input and spot failed requests before the next morning.

For people using it this way, how much extra effort does the asynchronous workflow add? I’m especially curious about handling partial failures and deciding whether the savings are worth giving up real-time results.