Context window length with further fine tuning and a good system prompt... Nothing too much different other than anthro correctly banked on larger context windows (and stability with training/eval) is the key to massive improvements.