How many tokens and which models were used?