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

We introduce GDPval, a benchmark evaluating AI model capabilities on real-world economically valuable tasks. GDPval covers the majority of U.S. Bureau of Labor Statistics Work Activities for 44 occupations [p 2 ] across the top 9 sectors contributing to U.S. GDP (Gross Domestic Product). Tasks are constructed from the representative work of industry professionals with an average of 14 years of experience . [p 3 ] We find that frontier model performance on GDPval is improving roughly linearly over time , [p 6 ] and that the current best frontier models are approaching industry experts in deliverable quality . [p 5 ] We analyze the potential for frontier models, when paired with human oversight, to perform GDPval tasks cheaper and faster than unaided experts. We also demonstrate that increased reasoning effort, increased task context, and increased scaffolding improves model performance on GDPval. Finally, we open-source a gold subset of 220 tasks and provide a public automated grading service at this http URL to facilitate future research in understanding real-world model capabilities.

The benchmark still covers a limited slice of knowledge work performed on a computer, and it does not establish labor-market impact. [p 9 ]

Claude Opus 4.1 was the best-performing model on the GDPval gold subset. [p 5 ] On the open 220-task gold subset, 47.6% of Claude Opus 4.1’s rated deliverables were judged better than or as good as the human expert deliverable. [p 6 ] This is a combined wins-and-ties figure, not wins alone. It applies to the best-performing model reported, not to all models or all occupations, and it does not mean the model performs as well as experts across all work.

Experts particularly favored Claude’s aesthetics and formatting, while GPT-5 stood out more for accuracy and instruction-following. [p 6 ] The attempted blinding was imperfect, because graders might infer model identity from stylistic tells. [p 5 ]

The paper models a “try n times, then fix it yourself” scenario. A person samples and reviews an output, retries if it falls short, and completes the task if the attempts fail. For GPT-5, using the model’s 39.0% win rate, this scenario estimates 1.39× faster and 1.63× cheaper than unaided expert completion. [p 13 ] These are estimated workflow ratios, not observed workplace time savings. The calculation includes model completion, expert review, and potential human rework. It excludes review time for human work and the cost of catastrophic mistakes, and the wage estimates may understate the market cost of the highly experienced experts recruited. [p 13 ]