AI agents failed to pass DAYJOB: Tax, a benchmark evaluating their ability to handle complex, real-world tax work, including navigating messy records, identifying consequential issues, and making professional judgments with incomplete information. The top-performing agent, Opus 5.5, achieved a 9.2% pass rate, while others, such as Astra and Claude Sonnet 5.5, performed significantly worse. AI summary
Firehose
Filtered to Hacker News, tagged “taxation” · clear filters
Browse: People · Companies · Papers · Podcasts · Hacker News · Deep dives