GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks
Integrity rank #19 of 65 · 11 models scored · top score 49.7 · GPT-5.2
Ability to produce expert-quality professional work products across 44 occupations
Strongest on harness comparability, weakest on contamination resistance.
Reported scores — protocols may differ. How we compare
| # | Model | Score | Evidence | Measured |
|---|---|---|---|---|
| 1 | GPT-5.2 | 49.7 | unverifiedT2 | 2025-12-11 |
| 2 | Claude Opus 4.5 | 45.5 | unverifiedT2 | 2025-11-24 |
| 3 | Claude Opus 4.1 | 43.6 | unverifiedT2 | 2025-08-05 |
| 4 | Claude Sonnet 4.5 | 42.5 | unverifiedT2 | 2025-09-29 |
| 5 | Gemini 3 Pro | 40.3 | unverifiedT2 | 2025-11-18 |
| 6 | GPT-5 | 34.8 | unverifiedT2 | 2025-08-07 |
| 7 | o3 | 30.8 | unverifiedT2 | 2025-04-16 |
| 8 | o4-mini | 25.3 | unverifiedT2 | 2025-04-16 |
| 9 | Gemini 2.5 Pro (Jun 2025) | 23.3 | unverifiedT2 | 2025-06-05 |
| 10 | Grok 4 | 21.1 | unverifiedT2 | 2025-07-09 |
| 11 | GPT-4o (Nov 2024) | 9.9 | unverifiedT2 | 2024-11-20 |
Continue from this leaderboard into the tightest head-to-head reads.
Weighted composite. Each dimension opens to the source-backed sub-signals it is computed from.
Claims drawn from cited facts, not live model generation.
This benchmark scores 11 models, and the spread between the best and worst of them is 39.8 points. That wide spread supports real separation between models, indicating that rank differences here are likely to reflect genuine capability gaps rather than close, ambiguous scores.
3 cited facts
This benchmark has an integrity score of 93, an integrity grade of A, and an integrity rank of 17 among 62 benchmarks in the universe. Its weakest integrity component is contamination, which narrows how much the top scores can be trusted. Together these figures describe the benchmark's health as a discriminator of frontier models under a disclosed harness, not any model's capability.
6 cited facts
This benchmark is not saturated: the best result stands at 49.7 points on a hundred-point scale, leaving 50.3 points of headroom to the ceiling. Only one model sits within a couple of points of that top score, so the top of the leaderboard is not crowded. Because the benchmark is not saturated, there is still genuine room to separate the field at the top, and differences between leaders can reflect real capability gaps rather than noise.
5 cited facts
The harness comparability label indicates a consistent harness, so scores should be read as task performance under that disclosed harness rather than deployed capability and compared only when that same harness is used. The test set privacy label is semi-private, meaning a public portion and a held-out portion, so high scores deserve more contamination scrutiny than a fully held-out set even though contamination history shows no documented contamination. Because harness notes indicate sensitivity to grader choice and scaffold, scores from different harnesses should not be compared head-to-head; any apparent advantage should be interpreted cautiously.
5 cited facts
This benchmark measures a model's ability to produce expert-quality professional work products across a wide range of occupations. Its test items were built from real deliverables created by experienced professionals and sampled across major economic sectors and occupations, so the suite aims to reflect genuine knowledge work rather than synthetic puzzles. The design rests on two assumptions: that blind human preference over one-shot deliverables stands in for real economic task value, and that sampling occupations by their share of economic output adequately represents knowledge work. The sharpest caveat is that these are one-shot deliverables rather than multi-day jobs, and preference grading is subjective and rater-dependent, so the released automated grader only approximates human judgment and the results speak only to white-collar knowledge work.
5 cited facts
GDPval (GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks): GDPval evaluates AI models on real-world, economically valuable deliverables (documents, slides, spreadsheets, diagrams, CAD) drawn from 44 occupations across the 9 largest U.S. GDP-contributing sectors; scored as the rate at which model outputs beat ('wins') or match ('ties') expert human deliverables under blind pairwise grading. Best model (Claude Opus 4.1) reached 47.6% wins+ties as of Sept 2025.
GPT-5.2 leads GDPval at 49.7. The full leaderboard above lists every recorded measurement, not just the headline number.
11 models have recorded scores on GDPval, spanning a score spread of 39.8.
tensor.news grades GDPval A for integrity (score 93/100), ranking #19 of 65 benchmarks we assess across discrimination, saturation, contamination resistance, harness comparability, and freshness.
No — GDPval still has headroom and continues to discriminate between models rather than bunching them at the ceiling.
Its test set is semi-private, which lowers — but does not eliminate — the risk that training data leaked into the questions.
Scores are reported under a consistent harness, so comparisons on GDPval are reasonably apples-to-apples.
Every GDPval measurement is source-backed and tagged reproduced, self-reported, or unverified; the underlying record cites arxiv.org/abs/2510.04374.