Senior SWE-Bench benchmark: which model API scores highest for agent coding