First scale, then enrich: How the right execution strategy helped us reach state-of-the-art on SWE-rebench

AI21 Labs achieved a new state-of-the-art result on the SWE-rebench benchmark with a 60.9% issue resolve rate for 123 issues between December and March, surpassing the previous best published result. The team reversed the conventional pipeline order by generating candidate solutions first through horizontal scaling and then using those candidates to guide targeted context extraction, improving the resolve rate from a 53.8% baseline ReAct loop while staying within the computational budget of leading competitors like Claude Code.