Patwardhan, Dias, Proehl et al. — OpenAI — 2025

GDPval: AI on Real-World Economic Tasks

Can frontier AI models do the actual work of industry professionals? A benchmark of 1,320 tasks from 44 occupations across 9 GDP sectors, graded by experts with 14 years average experience. The answer: they are getting close.

Prerequisites: LLM basics + Agent evaluation concepts
10
Chapters
5+
Simulations

Chapter 0: The Problem

You want to know whether AI can do economically valuable work. Not solve math puzzles. Not ace multiple-choice exams. Not write hello-world programs. You want to know: can a model write a legal brief that a partner at an AmLaw 100 firm would accept? Can it build a financial forecast that a 20-year Wall Street veteran would trust? Can it produce a marketing deck that a retail executive would actually present to her board?

Every existing AI benchmark fails this test. MMLU measures academic knowledge. HumanEval measures coding. GPQA measures graduate-level reasoning. These are all academic proxies for capability. They tell you the model is smart. They do not tell you the model is useful.

The core disconnect: AI benchmarks measure the kinds of tasks professors assign to students. Real economic value comes from the kinds of tasks employers assign to experienced professionals. A model that scores 95% on MMLU might still produce a legal brief that gets laughed out of court, a financial model with rookie errors in the discount rate, or a marketing deck with clip-art aesthetics. Academic smarts and professional competence are different things.

The problem runs deeper than task selection. Consider what makes professional work hard:

Historical precedent makes this urgent. Electricity took 40 years to reshape manufacturing. Computers took decades to show up in productivity statistics (the Solow Paradox). We cannot wait for lagging indicators like adoption rates or GDP growth to tell us whether AI is economically transformative. We need a leading indicator: directly measure what models can do on real work, right now.

Concept: Existing AI benchmarks are academic tests that do not measure economic value. Professional work demands subjectivity, multi-modality, long horizons, and domain expertise that no current benchmark captures. Realization: To know whether AI will transform the economy, we need to measure AI on the economy's actual tasks — not on proxies. GDPval is the first benchmark designed from this principle.
Why do high scores on academic benchmarks (MMLU, GPQA, HumanEval) fail to predict whether AI can do economically valuable work?

Chapter 1: The Key Insight

GDPval's design principle is deceptively simple: map tasks to actual GDP-contributing occupations, then have the professionals who do those jobs grade the outputs.

This is not "find some hard tasks and see if the model can do them." It is a systematic, top-down approach to covering the U.S. economy:

Step 1: Identify top GDP sectors
Select the 9 sectors that each contribute over 5% to U.S. GDP (as measured by Q2 2024 Value Added by Industry). Together these cover roughly 75% of the economy.
Step 2: Select top occupations per sector
Within each sector, pick the 5 occupations that contribute most to total wages and compensation, filtering for predominantly digital work. This yields 44 occupations earning $3 trillion annually.
Step 3: Recruit industry experts
Hire professionals with at least 4 years of experience (average 14 years) from top firms. They pass video interviews, background checks, training, and a qualification quiz. Less than 10% of applicants are accepted.
Step 4: Construct tasks from real work
Each expert creates tasks based on their actual professional work. Tasks are classified against the BLS O*NET taxonomy to ensure representative coverage of Work Activities. Average task takes 7 hours to complete.
Step 5: Grade by pairwise expert comparison
Different experts in the same occupation blindly compare model outputs against human expert outputs. No multiple-choice rubric — just professional judgment: which deliverable is better?

The brilliance of this design is in the grading. Most benchmarks grade against a fixed correct answer, which works for math but fails for subjective work. GDPval uses pairwise expert comparison: show a grader two unlabeled deliverables (one from a model, one from a human expert) and ask which is better. This mirrors how real work is evaluated — by peers and supervisors using professional judgment.

Key insight: The metric is win rate — how often the model's deliverable is rated better than the human expert's. This has no upper limit. As models improve, you can swap in stronger baselines. Unlike accuracy-on-a-test, win rate never saturates.
GDPval Pipeline

The five-step process from GDP sectors to graded model outputs. Click each stage to see details.

Stage 1: GDP Sectors
Why does GDPval use pairwise expert comparison instead of a fixed rubric or correct-answer check?

Chapter 2: Task Design

GDPval covers 1,320 tasks in the full set (at least 30 per occupation) and open-sources a 220-task gold subset (5 per occupation). The tasks span 44 occupations across 9 sectors. Let's look at what this actually covers.

The 9 Sectors

Sector% GDPTop Occupations (sample)
Real Estate & Leasing13.8%Property Managers, RE Agents, RE Brokers
Government11.3%Compliance Officers, Admin Managers, Social Workers
Manufacturing10.0%Production Supervisors, Industrial Engineers, Mechanical Engineers
Prof./Scientific/Tech8.1%Software Developers ($239B), Lawyers ($137B), Accountants ($135B)
Healthcare7.6%Registered Nurses ($323B), Medical Managers, Nurse Practitioners
Finance & Insurance7.4%Financial Managers ($148B), Customer Service Reps, Securities Agents
Retail Trade6.3%General/Operations Managers ($477B), Retail Supervisors, Pharmacists
Wholesale Trade5.8%Sales Reps ($103B), Sales Managers ($97B)
Information5.4%Producers/Directors, Editors, Journalists

Notice the compensation numbers. The Professional/Scientific/Technical sector alone has software developers earning $239 billion in aggregate compensation. These are not toy tasks — they represent the most economically valuable knowledge work in the United States.

How Tasks Were Built

Each task has two components: a request (the assignment, often with reference files like spreadsheets, PDFs, images, or audio) and a deliverable (the work product). Experts classified each task against O*NET occupational tasks to ensure coverage. The dataset covers 208 unique O*NET tasks, 25 skills, and 26 work activities.

Real-world complexity: 67.7% of tasks require interaction with reference files. Tasks use spreadsheets, CAD files, slide decks, photos, video, audio, and customer conversations. The gold subset has up to 17 reference files per task; the full set has up to 38. Average completion time is 7 hours for the gold subset, with some tasks spanning multiple weeks.

Task Statistics

MetricMeanMedianMax
Quality rating (1-5)4.474.505.00
Difficulty (1-5)3.323.005.00
Representativeness (1-5)4.505.005.00
Time to complete (hours)9.495.00100
Dollar value$398$175$4,114

Each task went through a multi-stage quality control pipeline: automated model screening, then at minimum three human expert reviews (averaging five reviews per task). Experts iteratively revised tasks between review stages. The result: 89% of tasks were rated as well-specified by occupational graders.

Sector Coverage Map

The 9 GDP sectors and their relative economic weight. Hover over sectors to see occupation count and compensation.

GDPval tasks average 7 hours of expert completion time. Why does this matter for evaluating AI economic impact?

Chapter 3: Grading

Grading is the hardest part of GDPval. You cannot auto-grade a legal brief the way you auto-grade a math problem. The paper uses two grading approaches, and the tension between them reveals something important about evaluating real work.

Human Expert Grading (Primary)

Experts in the relevant occupation are shown a request, reference files, and two or more unlabeled deliverables. They rank them and provide detailed justifications. The comparison is blinded — no model names, no labels. (Though stylistic tells remain: OpenAI models favor em dashes, Claude uses first-person phrasing, Grok occasionally refers to itself.)

On average, grading a single comparison takes over an hour. The grader must read the request, examine reference files, carefully review each deliverable, and write a justification. This is expensive: the average review costs $86 (based on median occupational wages times review time of 109 minutes).

Automated Grading (Experimental)

For the open-sourced gold subset, the team trained a grading model to mimic expert pairwise comparisons. It achieves 66% agreement with human experts — only 5% below human inter-rater agreement of 71%.

Key insight: Human experts only agree with each other 71% of the time. This is not a flaw — it reflects the genuine subjectivity of professional work. When an automated grader hits 66% agreement, it is remarkably close to the ceiling set by human disagreement itself.

Human Baseline

The human baseline deliverables were created by the same pool of industry experts. Average completion time: 404 minutes (~6.7 hours). Average cost per task: $361 (based on BLS median hourly wages). These are experienced professionals at top firms — Goldman Sachs, Apple, JPMorgan, Lockheed Martin, BBC News, and dozens more.

Grading Agreement

Human experts agree with each other 71% of the time. The automated grader agrees with humans 66% of the time. This 5% gap is small given the inherent subjectivity of professional work.

The automated grader agrees with human experts 66% of the time. Human experts agree with each other 71% of the time. What does this 5% gap tell us?

Chapter 4: Results Overview

Here is the headline result that makes GDPval matter: frontier models are approaching parity with industry experts on real-world economic tasks.

The paper evaluated seven models using blind pairwise comparison by professional experts: GPT-4o, o4-mini, o3, GPT-5, Claude Opus 4.1, Gemini 2.5 Pro, and Grok 4. The results:

ModelWin RateWin + Tie RateKey Strength
Claude Opus 4.1~32%47.6%Aesthetics (formatting, layout, PDFs, slides)
GPT-5 (high)~30%~45%Accuracy (instructions, calculations, text)
Grok 4~27%~42%Fewer accuracy errors
Gemini 2.5 Pro~25%~40%Broad capability
o3~24%~38%Reasoning-heavy tasks
o4-mini~20%~34%Speed/cost efficiency
GPT-4o~12.5%~25%Baseline comparison
The big number: Claude Opus 4.1 matched or beat expert humans on 47.6% of tasks. Nearly half the time, a model's deliverable was at least as good as what a 14-year industry veteran produced. This is not a hypothetical — these are real legal briefs, financial analyses, marketing decks, and engineering designs.

Just as striking: OpenAI frontier model performance improved roughly linearly over time. Not exponential. From GPT-4o to o4-mini to o3 to GPT-5, each generation gained roughly the same increment. This is a crucial finding — it means we can somewhat predict when models will reach full expert parity on GDPval tasks.

Where Models Fail

The paper built a clustering pipeline to analyze why experts preferred or rejected deliverables. The top failure modes across all models:

  1. Instruction-following failures — the most common loss reason for Claude, Grok, and Gemini. Models failed to fully follow the request.
  2. Formatting errors — GPT-5's most common failure. Black squares in PDFs, broken slide layouts, font issues.
  3. Hallucinated data — all models sometimes fabricated numbers or miscalculated.
  4. Missing deliverables — Gemini and Grok frequently promised files but failed to provide them.
  5. Ignored reference data — models sometimes didn't use the provided reference files.

When GPT-5 failures were categorized by severity: most were "acceptable but subpar" (the model produced usable work, just not as good as the human). About 29% were "bad" or "catastrophic," with only ~3% truly catastrophic (harmful or dangerously wrong).

Model Win Rates Over Time

Drag the slider to see how OpenAI frontier model performance improved roughly linearly. The dashed line at 50% represents expert parity.

GPT-5
Concept: Frontier models are approaching expert quality on real economic tasks, with the best models matching or beating experts nearly half the time. Realization: Linear improvement, not exponential, means full expert parity is years away on hard tasks. But the models are already good enough to be useful in a human-oversight workflow today.
Claude Opus 4.1 excels at aesthetics (PDFs, slides) while GPT-5 excels at accuracy (instructions, calculations). What does this suggest about evaluating AI on real work?

Chapter 5: Cost & Speed

Even when a model doesn't beat the human expert, it might still be economically useful. The question is not "is the model as good as a human?" but "is model + human oversight cheaper and faster than a human alone?"

The Numbers

For the gold subset, the paper computes three key quantities per task:

Three Scenarios

The paper analyzes increasingly realistic workflows:

Naive: Just Use the Model
Speed ratio: GPT-5 is 90x faster than a human. Cost ratio: 474x cheaper. But this ignores quality — you'd accept whatever the model produces.
Try Once, Then Fix
Use the model, have an expert review. If unsatisfactory, the expert does it themselves. GPT-5: 1.12x faster, 1.18x cheaper. Modest but real savings even accounting for review overhead and human fallback.
Try N Times, Then Fix
Resample from the model up to N times, reviewing each. Only fall back to human if all N fail. GPT-5: 1.39x faster, 1.63x cheaper. Each resample is cheap, and with a 39% win rate, you likely get an acceptable output within a few tries.
Key insight: The formula for "try N times" is elegant. Expected time for task i: E[Tn,i] = (MT + RT) · (1 - (1-w)n) / w + (1-w)n · HT. As N → ∞, time converges to (MT + RT) / w. Even a model with only 39% win rate becomes 1.39x faster than an unaided expert if you keep sampling.
ModelWin RateSpeed ImprovementCost Improvement
NaiveTry 1xTry NxNaiveTry 1xTry Nx
GPT-4o12.5%327x0.87x0.46x5172x0.90x0.53x
o4-mini29.1%186x1.02x1.06x1265x1.06x1.22x
o335.2%161x1.08x1.28x480x1.13x1.47x
GPT-539.0%90x1.12x1.39x474x1.18x1.63x

Notice the critical threshold: GPT-4o with 12.5% win rate actually makes workers slower (0.87x) and more expensive (0.90x) in the "try once" scenario. The review overhead eats any savings. But once win rate crosses roughly 25-30%, the economics flip positive. Models become genuinely helpful even with human oversight.

Cost-Speed Tradeoff

Adjust the model win rate to see how the "try N times" economics change. The break-even point is where the improvement ratio crosses 1.0x.

39%
GPT-4o (12.5% win rate) is 327x faster than a human in raw speed, but only 0.87x faster in the "try once" scenario. Why?

Chapter 6: Scaling Effects

Three levers improve model performance on GDPval: more reasoning, better prompts, and improved scaffolding. The paper tests each one.

Reasoning Effort

Running o3 and GPT-5 at low, medium, and high reasoning effort shows a clear trend: more thinking time improves quality. This is not surprising given the nature of the tasks — a 7-hour professional task benefits from deeper reasoning about what the client wants, how to structure the deliverable, and whether the output actually addresses the prompt.

Prompt Engineering

Many GPT-5 failures were obvious formatting errors — black squares in PDFs, broken slide layouts, non-standard Unicode characters. The team created a general prompt that instructs the model to:

Results were dramatic. Black-square artifacts: eliminated entirely (from affecting over half of PDFs to zero). Egregious formatting errors in PowerPoints: reduced from 86% to 64%. Models using their multi-modal capabilities to inspect deliverables: jumped from 15% to 97%. Overall win rate: +5 percentage points.

The lesson: These are easy performance gains. The models already have the capability to self-check their work — they just do not do it by default. A simple prompt telling the model "look at what you produced and fix obvious errors" recovers significant quality. This suggests current GDPval scores are a lower bound on what careful scaffolding can achieve.

Scaffolding and Best-of-N

The team also enabled GET requests in the container, performed best-of-4 sampling with a GPT-5 judge picking the best output, and combined these with the prompt improvements. Each layer stacked: reasoning effort + prompt engineering + scaffolding all contributed independently.

Context Sensitivity

In a separate experiment, they created under-contextualized versions of GDPval tasks — prompts that were only 42% the token length of the originals, omitting details about where to find data in reference files, how to approach the problem, and formatting expectations. Model performance degraded significantly. This reveals that much of professional skill is knowing how to figure out context — understanding what a client actually wants when the brief is vague.

Scaling Levers

Toggle each lever on/off to see its estimated effect on model performance. All three stack.

Prompting the model to visually inspect its own outputs increased self-inspection from 15% to 97% and eliminated black-square PDF artifacts entirely. What does this reveal?

Chapter 7: Sector Breakdown

The aggregate win rates mask enormous variation across sectors and occupations. Some sectors are nearly solved; others remain stubbornly hard.

Where AI Approaches Parity

Three sectors stand out where the strongest models approach expert-level performance:

Where AI Struggles

Some sectors show low win rates across all models:

By Deliverable Type

Performance varies dramatically by file format:

By Task Duration

Win rates are highest for shorter tasks (0-2 hours) and decline steadily as completion time increases. This makes intuitive sense: longer tasks require more planning, more context management, and more sustained coherence — all areas where models still struggle relative to humans.

Sector Performance Heatmap

Relative model performance across sectors. Warmer colors indicate higher win rates. Hover to see details.

Concept: AI performance is not uniform — it varies enormously by sector, occupation, deliverable type, and task duration. Realization: "AI will replace knowledge workers" is far too coarse. The correct question is always: which tasks, in which sectors, for which deliverable types? GDPval gives us the resolution to answer this precisely.
Win rates are highest for short tasks (0-2 hours) and decline with task duration. Why?

Chapter 8: Implications

"Approaching expert quality" sounds like a headline. But what does it actually mean for the economy and for workers?

What 47.6% Win+Tie Means

Nearly half the time, a model matches or beats a 14-year veteran. But this is an average across 220 tasks in 44 occupations. In some occupations it is much higher; in others, much lower. And even a 47.6% rate means the model still loses more than half the time. This is not "AI replaces experts" — it is "AI assists experts in a meaningful fraction of their work."

The Human-in-the-Loop Economics

The cost/speed analysis from Chapter 5 is the real economic story. Even with review overhead and human fallback:

The economic arithmetic: If AI saves 20% of expert time across $3 trillion in annual compensation, that is $600 billion per year in productivity gains — from just 44 digital knowledge work occupations. The actual economy has hundreds more occupations, non-digital work, and second-order effects. The total impact could be much larger.

What "Linear Improvement" Means

GDPval finds that OpenAI model performance improves roughly linearly over time, not exponentially. This is significant: linear improvement means full expert parity is predictable but not imminent. If models gain ~5-8 percentage points per generation and a generation takes 6-12 months, reaching 50%+ average win rate (true parity) is likely several years away for the hardest tasks.

But this also means the improvement is steady and reliable. Unlike exponential curves that might plateau suddenly, a linear trend suggests consistent engineering progress. Every generation makes AI slightly more useful for a slightly broader set of tasks.

Limitations to Remember

GDPval only covers self-contained, digital knowledge work tasks. It does not test:

These limitations mean GDPval likely overestimates AI capability for the messiest, most interactive real-world work, while accurately measuring capability on well-specified digital deliverables.

GDPval tasks are "precisely-specified and one-shot." Why does this limitation mean GDPval likely overestimates AI's real-world economic impact?

Chapter 9: Connections

GDPval sits at the intersection of AI evaluation and labor economics. Here are the key threads that connect it to the broader research landscape.

AI Evaluation

Labor Economics

Why GDPval Matters Going Forward

The open-sourced gold subset of 220 tasks with an automated grader at evals.openai.com means anyone can run their model against real economic tasks. As models improve, the benchmark does not saturate — you simply compare against increasingly strong baselines. The pairwise comparison design ensures there is always room to measure improvement.

Concept: GDPval bridges AI evaluation and labor economics, measuring what no prior work has: can AI do the actual work that drives GDP? Realization: With 47.6% win+tie rate today and linear improvement, the question is not whether AI will transform knowledge work, but how fast — and GDPval gives us the ruler to measure the pace.
Why is GDPval described as a "leading indicator" of AI's economic impact, unlike adoption rates or GDP statistics?