The GDP Benchmark: A New Frontier for Measuring AI Capabilities in Professional Knowledge Work, by Jonathan H. Westover PhD

Abstract: This article examines OpenAI's recently released GDPval benchmark, which represents a significant advancement in evaluating artificial intelligence capabilities on economically valuable knowledge work. Unlike previous AI evaluations that focus on academic reasoning or specific domains, GDPval assesses performance on real-world tasks spanning 44 occupations across 9 major economic sectors that contribute $3 trillion annually to the U.S. economy. Analysis of benchmark results reveals that frontier AI models are approaching expert-level performance on many professional tasks, with the best models winning or tying with human experts approximately 50% of the time. The benchmark also demonstrates that human-AI collaboration strategies can potentially increase productivity while maintaining quality. This article synthesizes the methodology, findings, and implications of GDPval, offering evidence-based recommendations for organizations seeking to integrate AI capabilities into knowledge work processes. While these results show impressive AI progress on standalone professional tasks, they should be interpreted as indicators of task-level capabilities rather than predictions of occupational displacement.