GDPval

Patwardhan et al., OpenAI, 2025

What it measures

A published test we don't mirror: it supplies a reference figure, not test items

GDPval measures AI on real work: tasks from 44 occupations across the nine sectors that contribute most to US GDP, written by professionals with 14 years of experience on average.

Expert graders compare a model's deliverable with one made by a professional and pick the better one. Two experts compared this way tie at 50%, which is why the human side sits at 50 by design.

Example question

Written for this page in the style of the test, not one of its questions, so the test set stays unspoiled

A property manager's noticespecimen · not scored

Produce the deliverable a professional in this occupation would hand in.

(a) Task

You manage a 40-unit residential building. A contractor will shut off the water on Tuesday from 9:00 to 13:00 to replace a pipe. Write a one-page notice for tenants: what is happening and when, how to prepare, and who to call. The contractor's schedule is attached.

(b) Grading

An expert in the same occupation sees this deliverable next to one written by a professional, unlabelled, and picks the better one or calls a tie.

Figure 1. A deliverable task. There is no answer key: quality is judged against human work. In the style of GDPval. Answer: None. The model scores when the grader prefers its notice or calls a tie.

Human vs AI

The faculty-level figures we take from it, 0–100

GenerationGap -34.9

Humans

50.0

GDPval, expert vs expert (parity by design)

Best published AI

84.9

GDPval wins or ties, GPT-5.5 (vendor-reported)

These are third-party results on different questions from ours. Where the numbers come from

All reference tests →