GDPval, One Year On: What Testing AI Against 14-Year Veterans Actually Showed

Most AI benchmarks look like exams. GDPval looks like an inbox. Instead of trivia or coding puzzles, it asks AI models to produce the kinds of deliverables real professionals hand in: a legal brief, an engineering concept, a nursing care plan, a customer support reply.

OpenAI introduced it in September 2025 and published the full paper in October 2025. A year on, it is still one of the clearest public tests of a practical question every team deploying AI agents faces: which work can you actually hand over?

How the test was built

GDPval covers 44 occupations across the 9 U.S. industries that each contribute more than 5% of GDP. Within each industry, OpenAI picked the occupations that pay the most in total wages and are mostly knowledge work, using Bureau of Labor Statistics and O*NET data.

The tasks were written by working professionals who averaged 14 years of experience. Each task went through an average of five rounds of expert review. The full set has 1,320 tasks, and a "gold" set of 220 tasks is open-sourced.

Importantly, the tasks are not one-line prompts. They come with reference files and context, and the expected outputs are documents, slides, spreadsheets, diagrams, and media.

Grading was blind. Experts from the same occupations compared AI deliverables with the human expert's own deliverable, without knowing which was which, and rated each AI output as better, as good as, or worse.

What it found

On the 220-task gold set, OpenAI reported that:

• The best model in the original comparison, Claude Opus 4.1, produced work rated as good as or better than the human expert's in just under half the tasks. It stood out on presentation and formatting.
• GPT-5 stood out on accuracy, such as finding domain-specific knowledge.
• Performance more than doubled from GPT-4o to GPT-5, roughly a year apart, following what OpenAI describes as a clear linear trend.
• Frontier models finished tasks roughly 100 times faster and 100 times cheaper than the experts.

That last number comes with a caveat OpenAI states plainly: it counts only model run time and API cost. It does not include the human oversight, iteration, and integration that real workplaces need.

What it did not test

The limitations section is where GDPval becomes most useful for anyone planning an agent rollout. OpenAI says the current version:

• is one-shot, so it does not capture revising a draft after feedback or spotting an anomaly and redoing an analysis;
• gives the model a clearly defined task with the files attached, whereas real work often starts with ambiguity, like a lawyer needing to talk to a client before deciding a brief is even the right deliverable.

The paper also reports that more reasoning effort, more task context, and more scaffolding each improved results.

Five lessons for delegating work to AI agents

Read as a guide to delegation rather than a scoreboard, GDPval points to a few practical rules.

1. Delegate deliverables, not jobs. The test measures tasks, and OpenAI notes that most jobs are more than a list of tasks. Pick specific, repeatable outputs to hand to an agent rather than a whole role.

2. Context is part of the task. GDPval tasks came with reference files, and adding context measurably helped. An agent that can reach the right documents, systems, and past examples will do better than a smart model working blind.

3. Plan for review, not just generation. "As good as an expert just under half the time" also means the rest of the time it was not. The 100x speed advantage only holds if checking the work stays cheap. Build the review step in from the start.

4. Try the agent first on well-specified work. OpenAI's own suggestion is that, on tasks where models are strong, giving the task to a model before a person would save time and money. That works best where the brief is clear and the output can be checked against a standard.

5. Keep ambiguity with people. The test deliberately removed the messy front end of real work. Deciding what the client actually needs is still a human step.

What this means for teams deploying AI agents going forward

GDPval did something useful by grounding the "can AI do my job" debate in real deliverables graded by real experts. The honest reading is neither hype nor dismissal. On clear, well-documented tasks, current models are already close to experienced professionals and far faster. On the parts of work the benchmark left out, iteration and ambiguity, the evidence is thinner.

For teams, that suggests a simple operating model: give agents the well-specified deliverables, give them the context a new colleague would need, and keep a person responsible for the brief at the start and the check at the end.

FAQ

What is GDPval? GDPval is an OpenAI benchmark that tests AI models on real work deliverables from 44 occupations across the 9 largest sectors of the U.S. economy, graded blind by industry experts.

How close are AI models to human experts on GDPval? In OpenAI's original results, the best model was rated as good as or better than experts in just under half of the 220 gold-set tasks.

What are GDPval's limits? OpenAI says it is one-shot and uses clearly defined tasks, so it does not measure revising work after feedback or handling ambiguous requests.