The best AI in the world scores 12% at real work

SemiAnalysis published GDPx on November 2, a 206-task benchmark that drops AI agents into realistic work environments across eleven domains instead of handing them a prompt and a document. The best model in the world, Anthropic's Claude Opus 5.5, completed just 12.0% of tasks, ahead of Google's Gemini 4 Carbon at 10.9% and OpenAI's GPT-6 Astra at 9.6%.

Every model scored zero on finance and legal tasks under binary grading, and the authors admit safety classifiers sandbagged Opus 5.5 and GPT-6 Astra where models refused rather than attempted. The benchmark is the industry's cold shower: the trillion-dollar automation pitch, measured against actual work, tops out at one task in eight.