Surge AI's DAYJOB: Tax benchmark finds the best agent passes only 10.8% of professional tax assignments
Surge AI released DAYJOB: Tax, 50 expert-reviewed agent tasks across eight practice areas, backed by nearly 1,800 client workspace files. A task passes only if the agent meets 95% of the rubric. Claude Sonnet 5.5 led with 10.8%, GPT-6 Astra reached 4.8%, and failures were mostly judgment errors rather than math errors. Only two tasks are public.
Why it matters
A grounding check on agent claims for regulated professional work. Even top models rarely produce a deliverable that holds up under review.
Source: Read the full story on Surge AI