Google trails SWE-bench's leader Anthropic by 1.0 points; 9 labs are scored
Verified · bash-only mini-SWE-agent (same scaffold for every model) · 1 attempt(s), each company's best model in that set-up · one of the AI Trajectory page's inputs; the page's 1-10 scores are a rule-based reading and stay there
AI · AI · SWE-bench · 7 Nov 2026