I've resorted to using concurrent (not batch) requests for now, even though it's twice as expensive, because at least I can reliably get output. It's also a lot slower because the batch requests are processed in parallel