I've not seen anyone seriously attempting to benchmark chatgpt output, without heavily cherry picking it first.