AI matched or beat professionals on 84.9% of sample work tasks by April 2026, up from 47.6%
Experts blind-graded AI and professional work, such as legal briefs, from 44 jobs. The 84.9% is OpenAI's own count on its own test.
OpenAI introduced GDPval in September 2025 to test AI on real work rather than exam questions. It covers 44 knowledge-work jobs in the 9 sectors that contribute most to US GDP (the size of the economy), with 1,320 tasks, 220 of them in a public “gold” set. Professionals with an average of 14 years’ experience wrote the tasks. The work products include legal briefs, spreadsheets, slide decks and care plans. Expert graders compared AI and human work without knowing which was which.
At launch, the best model was a competitor’s: Claude Opus 4.1’s work was rated as good as or better than the expert’s on 47.6% of gold-set tasks. GPT-5 (high) scored 38.8% and GPT-4o 12.4%. OpenAI also said models finished tasks about 100 times faster and 100 times cheaper, counting only computing time and API cost (the fee for using the model).
Every later figure is OpenAI’s own. GPT-5.2 Thinking reached 70.9% in December 2025, which OpenAI called its first model at or above human expert level. In April 2026, GPT-5.5 scored 84.9%, GPT-5.4 83.0%, Claude Opus 4.7 80.3% and Gemini 3.1 Pro 67.3%.
What it means
On clear, self-contained tasks, the best AI now produces work that experts rate as good as a professional’s most of the time. That matters for jobs built around documents like these. It does not show AI can do a whole job, which also means gathering context, revising and working with others, and the figures come from the company that built the test.
Caveats
- OpenAI built, ran and reports the benchmark; every figure after launch is OpenAI’s own evaluation.
- Tasks are one-shot and clearly specified, so they leave out gathering context, iterating and redrafting. Ties count toward the headline figure.
- The speed and cost comparisons leave out the human oversight, iteration and integration real workplaces need.
- Matching experts on sample tasks isn’t the same as automating a whole job.
- Newer models are tested on a separate third-party version (Artificial Analysis’s GDPval-AA v2); no verified GDPval figure exists for models after GPT-5.5.