Best AI agent passed under 11% of real tax assignments in Surge test

Surge AI released DAYJOB: Tax, a benchmark of 50 US tax assignments built and reviewed by tax professionals, with 1,789 files including returns, ledgers, receipts, IRS notices and emails. Agents have to spot issues the client never asked about, not just do the maths, and an attempt passes only if it meets at least 95% of the grading criteria. The top score was Claude Sonnet 5.5 at 10.8%, ahead of Claude Fable 5.1 (9.6%) and Claude Opus 5.5 (9.2%). In one example, Astra missed a deadline that could cost a home seller a tax break in all five attempts, while Opus caught it every time. The full benchmark is available only on request, so the results have not been independently reproduced. On Hacker News (193 points), of 24 top-level comments, 3 said they use models to check their own returns and found real errors or savings, 3 reported mixed or poor results or doubted models can handle tax reliably yet, 4 commenters questioned the benchmark itself (no human baseline, or agents given no tax-specific tools), and 1 raised privacy. No consensus. Sorted take: models can help check work, but this shows they still miss the issues a client never mentions. Use them as a second reviewer, not the preparer.