“Achieving a 90-95% score on the SWE-Bench benchmark would represent an AI model's ability to autonomously perform a significant fraction of software engineering tasks.”