SWE-bench creator Ofir Press says sandbox environments and data cleaning drive 40% of recent AI coding gains

Custom human-made evaluation tasks can cost up to $500,000.