GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks

Key findings

1. 01
   Claude Opus 4.1's deliverables were rated better than or as good as the human expert's on 47.6% of the 220 gold-subset tasks.Graders were occupational experts comparing AI and human work blind. Claude did best on formatting and presentation; GPT-5 did best on accuracy. (p. 6)
2. 02
   Among OpenAI models, the win rate against experts rose from 12.5% for GPT-4o to 39.0% for GPT-5.OpenAI describes the improvement across its model releases as roughly linear over time. (p. 15)
3. 03
   When an expert tries GPT-5, reviews the output and redoes the work if it falls short, the process was estimated to be 1.12x to 1.39x faster and 1.18x to 1.63x cheaper than the expert alone.Raw model inference is roughly 90x faster than the expert, but review time and failed attempts erase most of that gap. The analysis does not count the cost of catastrophic mistakes. (p. 15)
4. 04
   About 29% of expert ratings of GPT-5's losing deliverables called them bad or catastrophic, with about 3% catastrophic.The most common rating for a GPT-5 failure was 'acceptable but subpar'. (p. 15)
5. 05
   The 44 occupations come from the 9 sectors contributing most to U.S. GDP and collectively earn $3 trillion a year.Tasks are one-shot and fully specified, with the context supplied up front, rather than interactive work that unfolds over days. (p. 2)