GDPval
In plain wordsAn OpenAI test that pits AI against seasoned professionals on genuine workplace assignments, from reports and spreadsheets to slide decks, in 44 occupations important to the US economy.
A benchmark of AI performance on paid knowledge work rather than exam questions, made of 1,320 tasks from 44 occupations, which were chosen from the nine industries that add most to US gross domestic product. Industry veterans, averaging 14 years in their fields, built each task from a genuine piece of professional work, often with reference files attached; the deliverables range from documents, slides and spreadsheets to diagrams and multimedia. Grading is by blind pairwise comparison: experts from the same occupation judge whether the model's deliverable is better than, as good as, or worse than a human professional's, and the headline metric is the share of tasks the model wins or ties. A 220-task gold subset and an experimental automated grader are public. In the September 2025 release, the best model (Claude Opus 4.1) was rated as good as or better than the experts on just under half of the tasks.
On GDPval the model wins or ties against professionals on most spreadsheet tasks but rarely on slide decks.
Origin
Introduced by OpenAI on 25 September 2025, with a paper led by Tejal Patwardhan; the name joins GDP (gross domestic product) and 'val' for evaluation.
Usage note
Tasks are one-shot and well specified, so GDPval omits much of real work: iteration, gathering context, coordination and accountability. OpenAI's claim that models were about 100 times faster and cheaper counted only inference time and API cost, not human oversight.
Related
- task exposuretask exposure, n.: A measure of how much of a job's work AI could, in principle, do or speed up. High exposure means many of the job's tasks are within AI's reach, not that the job will vanish.
- technological unemploymenttechnological unemployment, n.: Joblessness caused by machines or software taking over work that people used to be paid for, faster than new kinds of work appear for them.
- win ratewin rate, n.: The share of head-to-head comparisons in which judges — people or another AI — prefer one model's answer over a rival's for the same request.
- human evaluationhuman evaluation, n.: Judging an AI's outputs by asking people to rate or compare them, because for many tasks — writing, conversation, translation — no automatic score captures quality well.
- Remote Labor IndexRemote Labor Index, prop. n.: A test that gives AI agents real paid freelance projects — design, video, data work and more — and checks how many they finish to a standard a client would accept. Early agents managed very few.
- automationautomation, n.: Using machines or software to do work that people would otherwise do, from factory robots to programs that sort email or write reports.
Sources
- Patwardhan, T. et al. (2025). GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks. arXiv:2510.04374. arxiv.org/abs/2510.04374
- OpenAI (2025). Measuring the performance of our models on real-world tasks. openai.com/index/gdpval
Mentioned in
task exposuretask exposure, n.: A measure of how much of a job's work AI could, in principle, do or speed up. High exposure means many of the job's tasks are within AI's reach, not that the job will vanish., win ratewin rate, n.: The share of head-to-head comparisons in which judges — people or another AI — prefer one model's answer over a rival's for the same request., Remote Labor IndexRemote Labor Index, prop. n.: A test that gives AI agents real paid freelance projects — design, video, data work and more — and checks how many they finish to a standard a client would accept. Early agents managed very few.