If you can do it offline, batching solves this. We used a test dataset from CFPB and at n=20, it was 1.6x faster and 1.2x more expensive with no statistically meaningful accuracy dropoff. Did not tune for batch size, but it's possible that we could get to 30 and see better perf.