OpenAI Unveils GDPval Benchmark to Measure AI Against Real Work

OpenAI on Thursday published GDPval, a new benchmark designed to measure how advanced AI models perform on real, economically valuable work. The evaluation covers 1,320 tasks drawn from 44 occupations across nine industries that together account for a large share of U.S. economic activity, and tests whether models can produce deliverables such as legal briefs, engineering blueprints, presentations, spreadsheets, and nursing care plans in ways comparable to experienced professionals.

OpenAI said the tasks were authored and vetted by practitioners with an average of 14 years of experience and were selected using U.S. Bureau of Labor Statistics and O Net data to reflect common knowledge work. Rather than simple text prompts, GDPval supplies reference files and context and asks models to create multimodal outputs. Human graders performed blind comparisons of AI outputs and human work, and OpenAI supplemented that process with an experimental autograder it plans to release for research use while cautioning it is not as reliable as human evaluation. "This realism makes GDPval a more realistic test of how models might support professionals," the company wrote.