Surge’s DayJob Tax evaluates agents on 50 expert-created tax tasks containing 1,789 files. Under a rule requiring at least 95% of rubric criteria to pass a task, the leading model, Claude Sonnet 5.5, achieved a single-attempt pass rate of 10.8%. Reported failure modes included discovering unstated deadlines and reconciling conflicting documents. The result does not mean that all tax answers were only 10.8% accurate; it illustrates the need for human review within the scope of this vendor-created evaluation.