GDPval

Most benchmarks ask a model to answer a test question. GDPval asks something closer to what an employer would ask: can the model produce a deliverable — a legal brief, an engineering drawing, a nursing care plan, a customer-support transcript — that a professional in that field would accept as good work? OpenAI built the tasks from 44 occupations across the nine sectors that contribute most to US GDP, each one drafted and checked by an industry professional with an average of 14 years’ experience, so the test items resemble real work products rather than academic exam questions.

Grading is blind: human experts compare a model’s output against an actual human-produced answer to the same task, without being told which is which. At launch, OpenAI reported that GPT-5 and Claude Opus 4.1 — the top two systems it tested — produced work rated as equal to or better than the human comparison on close to half of tasks, and that performance had been improving steadily across model generations. Only a 220-task “gold” subset was released publicly, graded through an automated evaluation service OpenAI hosts, rather than the full 1,320-task set.

GDPval arrived as part of a broader shift, visible the same month in Scale AI’s SWE-bench Pro, away from saturating academic tests and toward harder evaluations grounded in professional output. Its explicit framing around GDP-weighted occupations rather than a research subfield made it a reference point in debates about AI’s effect on knowledge work specifically, and third-party trackers such as Artificial Analysis have since built leaderboards on top of the public gold subset as newer models are released.