Claude Sonnet 5.5 records 10.8% pass@1 on DAYJOB: Tax across 50 assignments

The benchmark supplies agents with 1,789 workspace files including 1,487 PDFs, 117 Excel workbooks, K-1s, IRS notices, and handwritten receipts. Agents must produce returns, workpapers, or advice while identifying unstated issues such as the August 2027 residence deadline that five Astra runs missed and five Opus runs caught. Eight practice areas range from individual filings to international transactions and disputes. Pass rates remain below 11% for all tested models. Sonnet 5.5 leads, followed by Fable 5.1 at 9.6% and Opus 5.5 at 9.2%. The Pareto frontier shows these models deliver the measured performance at higher per-trial cost than smaller variants that score under 5%. Prior DAYJOB releases on healthcare and finance displayed similar ceilings when agents confronted incomplete instructions and conflicting records. Tax work adds stricter requirements for recognizing consequential omissions that clients do not surface. Current results indicate that deployment in advisory or compliance roles will require additional retrieval, verification, and escalation layers before production use. Enterprise tax platforms already integrate model calls for routine extraction; extending them to judgment tasks will depend on measurable lifts above the 10.8% baseline rather than marketing claims.