I would also like to hear how you benchmarked the accuracy.

Comparing output between competing GPT models?