The lowest performing gpt-4 model is a fairly low bar now

i put this at 70% because gpt-4-0613 had very weak performance

That's an inclusive or, right? It only has to beat any GPT-4 on either MMLU or LMSYS, not both?

@IsaacCarruthers Correct

the minimum of gpt-4 or maximum? is beating gpt-4-0613 enough?

@CampbellHutcheson "any GPT4 series variant"

beating any GPT-4 variant on lmsys isn't that hard these days. Claude Haiku beats gpt-4-0613