Labor Vector
Labor Vector meta-aggregator / Labor performed with AI

GDPval

Labor performed with AILabor producing AIbenchmarkactive
GDPval: a benchmark (an evaluation dataset with an experimental automated grader). Operator: OpenAI: the paper's 19 authors (Tejal Patwardhan first; nine marked equal contribution) give OpenAI as their affiliation, with correspondence to an openai.com address; OpenAI's own post at this address answered 403 and is unread. Its unit: a task: a request (often with reference files) and a deliverable, constructed from actual work product of an expert professional; graded by blinded pairwise expert comparison of model and human deliverables (win rate). Whose labor: the work product of recruited industry professionals in 44 predominantly digital occupations across the 9 sectors contributing over 5% of U.S. GDP, recast as tasks for models; the same professionals' labor as task writers, reviewers and graders.
Visit GDPval ↗The /non rowThe DSL audit
We introduce GDPval, a benchmark evaluating AI model capabilities on real-world economically valuable tasks.

In its own words: https://arxiv.org/abs/2510.04374, read 2026-10-29

Who supplies, who directs, who gains

Supplies the laborrecruited industry professionals (minimum 4 years' experience, average 14) wrote the tasks and their deliverables, reviewed them (an average of five reviews per task) and graded model deliverables against human ones; the paper states they were 'well compensated' and gives no rate; it also thanks 'vendor partners' without naming them
Directs itOpenAI's research team: occupations chosen by GDP share and total wages; researchers reviewed all tasks, promoted experts to reviewers and ran quality-control checks; OpenAI models screened submissions, with experts retaining final responsibility for task accuracy
Receives the gains, bears the costsnot recorded as a distribution; the paper models time and cost savings for an expert who uses a model and fixes or redoes unsatisfactory output, valued at BLS median wages, and states those wages likely underestimate the experts' market cost

Each as the resource itself records it. Where it records nothing, the census says so.

What it records

IntersectionsLabor performed with AI: assistance, delegation, task fidelity, correction burdens and user labor drag; Labor producing AI: annotation, moderation, evaluation, maintenance, and creative or scholarly work entering training and retrieval
Kind of evidencetask benchmark
Unit of recorda task: a request (often with reference files) and a deliverable, constructed from actual work product of an expert professional; graded by blinded pairwise expert comparison of model and human deliverables (win rate) op
Whose laborthe work product of recruited industry professionals in 44 predominantly digital occupations across the 9 sectors contributing over 5% of U.S. GDP, recast as tasks for models; the same professionals' labor as task writers, reviewers and graders
Methodoccupations selected from BLS and O*NET data (top 5 by total wages per sector, predominantly digital by a GPT-4o task classification at a 60% threshold); expert-written tasks classified against O*NET tasks; multi-stage model and human review; blinded pairwise grading by occupational experts; an automated grader based on GPT-5-high for the gold subset
Coveragegeography: U.S. occupations and sectors (BLS, O*NET, OEWS May 2024 wages, Q2 2024 GDP shares); period: the first version, arXiv v1 of 5 October 2025 op
What it findsperformance of OpenAI frontier models increased roughly linearly over time on the gold subset; models are beginning to approach parity with industry experts: 47.6% of Claude Opus 4.1 deliverables were graded better than or as good as the human deliverable; the text also says model deliverables outperformed or matched experts' in just over half the tasks; Claude Opus 4.1 was best overall, strongest on aesthetics; GPT-5 strongest on accuracy; experts most often preferred the human deliverable because models failed to fully follow instructions; with expert review and redo counted, the payoff shrinks: GPT-5 at 39.0% win rate gives 90x naive speed but 1.12x (try once) and 1.39x (try n times); of GPT-5's losses, roughly 29% were rated bad or catastrophic (roughly 3% catastrophic); win rates decline as task completion time increases; reasoning effort, prompting and scaffolding raise performance; OpenAI's official GDPval leaderboard, as Inspect Evals (third party) reports it: 31.9% (wins only) and 34.6% (wins and ties), lower than the 47.3% Inspect Evals obtained for GPT-5 with thinking set to low through OpenAI's automated grader
Its stated limitsdataset size: a limited, initial cut of knowledge work tasks; self-contained computer-based knowledge work only: manual labor, physical tasks, tacit knowledge, proprietary tools and communication between individuals are out of scope; tasks are precisely specified, one-shot and non-interactive; the automated grader has limitations, and agrees less with experts on capable OpenAI models' outputs; cost: industry expert grading is expensive; the speed and cost analysis omits review of human deliverables and the cost of catastrophic mistakes; completion times are experts' self-reports; grading could not be fully blind

op the resource's own statement · 3p a third-party account · obs our observation · rd our reading. Hover a tag for its quote; each opens below with its source.

What this evidence can settle

Can answer

  • how frontier models' deliverables on expert-written, precisely specified, one-shot occupational tasks compare with experts' deliverables, as judged by blinded expert graders
  • how that comparison varies by model, sector, occupation, file type and task length
  • under its stated scenarios, how much expert time and cost a model could save once review and redo are counted

Cannot answer

  • how work is performed in actual workplaces, where context must be found and tasks are interactive
  • employment or wage outcomes
  • who receives the time and cost savings it models
  • physical, interpersonal or tacit-knowledge work
  • the terms and rate at which the professionals were paid

In the worked questions

Where it meets Labor Vector

rd The archive's reading: GDPval grades a model's deliverable against an expert's deliverable on a task an expert wrote, a fixed commission, which meets DSL at the question of whether the system's labor fulfils the commissioned task, measured in its own terms of win rate. Its 'try n times, then fix it yourself' accounting counts the expert's review and redo time against the model, which is where it meets user labor drag; task-origin is the experts' own work product, as the paper states it.

Entity

Canonical nameGDPval
Also known asGDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks, GDPval gold subset, GDPval full set
Addresshttps://openai.com/index/gdpval/
OperatorOpenAI: the paper's 19 authors (Tejal Patwardhan first; nine marked equal contribution) give OpenAI as their affiliation, with correspondence to an openai.com address; OpenAI's own post at this address answered 403 and is unread op
Founded2025-09-25 (the paper's PDF on OpenAI's CDN, by its creation date); arXiv v1 2025-10-05
Kindbenchmark (an evaluation dataset with an experimental automated grader)
Lifecycleactive — the first version (arXiv v1, 5 October 2025); the paper states the dataset is being expanded and that future iterations will add breadth, realism, interactivity and context; later releases unrecorded because OpenAI's post is unread op
Count1,320 tasks in the full set (30 per occupation) across 44 occupations and 9 sectors that collectively earn $3T annually; a gold subset of 220 tasks (5 per occupation) open-sourced, served on Hugging Face as 220 rows; the gold subset covers 208 unique O*NET tasks; Wikipedia (third party) gives the held-out set as 1,100 tasks available only internally in OpenAI (as of arXiv v1, 2025-10-05; the dataset page and Wikipedia read 2026-10-29) op
Data accessprompts and reference files of the 220-task gold subset open-sourced, scrubbed of information identifying the task's expert; the Hugging Face dataset (openai/gdpval) also serves each task's deliverable files and its rubric (rubric_pretty, rubric_json; rubric items marked author_type human); an experimental automated grader at evals.openai.com, which Inspect Evals (third party) describes as free and reached by uploading results to Hugging Face and submitting OpenAI's form; automated grading results not provided for 12 of the 220 tasks; the 1,100 held-out tasks only internal to OpenAI and closed to external models' grading (Wikipedia, third party) op
Licencethe paper: CC BY 4.0 (arXiv); the dataset page states no license op
Versions and datescdn pdf created: 2025-09-25; arxiv v1: 2025-10-05; read: 2026-10-29
Observation2026-10-29: read

Relations

Neighbours

Captures: the composition layer's account

The /non row

Evidenceevery field's quote, its source and the hash of the saved copy65
  • self description op
    We introduce GDPval, a benchmark evaluating AI model capabilities on real-world economically valuable tasks.
    https://arxiv.org/abs/2510.04374 · read 2026-10-29 · saved copy sha256 904393337a8b91ee…
  • operator op
    Amelia Glaese Jerry Tworek OpenAI
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • correspondence op
    Equal contribution. Correspondence to tejal@openai.com.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • post unread obs
    the post's address in sources/manifest.json; the page answered 403 and was not read
    https://openai.com/index/gdpval/ · read 2026-10-29 · saved copy sha256 b0a7ee49cca701cb…
  • authors op
    View a PDF of the paper titled GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks, by Tejal Patwardhan and 18 other authors
    https://arxiv.org/abs/2510.04374 · read 2026-10-29 · saved copy sha256 904393337a8b91ee…
  • positioning op
    Therefore, while informative when available, these methods are lagging indicators of AI impacts. We consider an alternate method for understanding the potential economic impacts of AI: directly measuring AI model capabilities.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • coverage op
    GDPval covers the majority of U.S. Bureau of Labor Statistics Work Activities for 44 occupations across the top 9 sectors contributing to U.S. GDP (Gross Domestic Product).
    https://arxiv.org/abs/2510.04374 · read 2026-10-29 · saved copy sha256 904393337a8b91ee…
  • count op
    the GDPval full set covers 1,320 tasks across 44 occupations, sourced to cover the majority of Work Activities tracked by O*NET for each occupation
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • count per occupation op
    with at least 30 tasks per occupation in the full set (and 5 tasks per occupation in the gold subset), across 44 occupations.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • count wages op
    GDPval covers tasks from 9 sectors and 44 occupations that collectively earn $3T annually.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • count onet op
    The dataset covered 208 unique O*NET tasks, 25 occupational skills, and 26 work activities.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • selection sectors op
    Selected sectors that contribute over 5% to US GDP as determined by Q2 2024 Value Added by Industry as a Percentage of Gross Domestic Product
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • selection digital op
    we prompted GPT-4o to classify each task as digital or non-digital, and then classified the overall occupation as digital if at least 60% of its component tasks were digital.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • unit op
    Each GDPval task consists of two primary components: a request (often with reference files) and a deliverable (work product).
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • task origin op
    Each task is constructed based on actual work product created by an expert professional.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • supplies op
    Tasks are constructed from the representative work of industry professionals with an average of 14 years of experience.
    https://arxiv.org/abs/2510.04374 · read 2026-10-29 · saved copy sha256 904393337a8b91ee…
  • recruitment op
    Experts were required to have a minimum of 4 years of professional experience in their occupation and a strong resume with a demonstrated history of professional recognition, promotion, and management responsibilities. The average expert had 14 years of experience. We further required experts to pass a video interview, a background check, a training and a quiz to participate in the project. Experts were well compensated for their time and experience.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • recruitment selectivity op
    Less than 10% of applicants were selected to contribute tasks to our full set.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • vendors op
    We are also grateful to our vendor partners for their collaboration and support throughout this research, and extend a special thank you to the industry experts who contributed their time and expertise to GDPval
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • review op
    All 1,320 tasks in the full GDPval set went through an iterative review pipeline involving both automated model-based screening and multiple stages of human expert review. Each task received an average of five human reviews (with a minimum of three reviews).
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • directs research team op
    Initially, our researchers manually reviewed all tasks to identify experts who produced consistently high-quality tasks; these individuals were trained and promoted to reviewers.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • directs models op
    Experts retained final responsibility for task accuracy and completeness; the model did not autonomously alter tasks.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • grading op
    we conducted blinded expert pairwise comparisons, where experts in the relevant occupation were presented with a request and reference files and asked to rank two or more unlabeled work deliverables.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • grading time op
    On average, grading each comparison for the gold subset took over an hour.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • automated grader op
    achieves 66% agreement with human expert graders, only 5% below human expert inter-rating agreement of 71%.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • task value op
    Each task’s dollar value was estimated by multiplying the average estimated completion time by median hourly wages for the corresponding occupation from OEWS data
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • task length op
    Tasks require an average of 7 hours of work for an expert professional to complete. On the high end, tasks span up to multiple weeks of work.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding trend op
    Figure 6: Performance of OpenAI frontier models increased roughly linearly over time on the GDPval gold subset.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding parity op
    Figure 5: On human pairwise comparisons, models are beginning to approach parity with industry experts on the GDPval gold subset.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding claude op
    on the GDPval gold subset, 47.6% of deliverables by Claude Opus 4.1 were graded as better than (wins) or as good as (ties) the human deliverable. Model deliverables outperformed or matched expert humans’ deliverables in just over half the tasks.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding strengths op
    Claude Opus 4.1 was the best performing model on the GDPval gold subset, excelling in particular on aesthetics (e.g., document formatting, slide layout), while GPT-5 excelled in particular on accuracy (e.g., carefully following instructions, performing correct calculations)
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding instructions op
    Figure 8: Across models, experts most often preferred the human deliverable because models failed to fully follow instructions on GDPval tasks.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding review redo op
    When incorporating time to review and redo work, the payoff from using a model shrinks.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding gpt5 ratios op
    gpt-5 39.0% 90x 1.12x 1.39x 474x 1.18x 1.63x
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • fix it setup op
    Here, an expert human samples from a model, reviews outputs, and if unsatisfactory, resamples and repeats. If no satisfactory output is obtained, the human completes the task themselves.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding failures op
    Another roughly 29% of ratings were for bad or catastrophic (with roughly 3% of failures marked as catastrophic).
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding length op
    Win rates are highest for shorter tasks (0–2 hours) and decline steadily as completion time increases.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • finding scaffolding op
    We also demonstrate that increased reasoning effort, increased task context, and increased scaffolding improves model performance on GDPval.
    https://arxiv.org/abs/2510.04374 · read 2026-10-29 · saved copy sha256 904393337a8b91ee…
  • gains wage basis op
    Because our experts were recruited specifically for being highly experienced in their field, these wage estimates likely underestimate their true market cost.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit size op
    The GDPval full set currently consists of only 44 occupations and 30 total tasks per occupation. It is therefore a limited, initial cut of knowledge work tasks, not a comprehensive evaluation of all possible occupational tasks. We are expanding the dataset size.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit scope op
    Manual labor and physical tasks are not included in the current version. Moreover, tasks that involve extensive tacit knowledge, access to personally identifiable information, use of proprietary software tools, or communication between individuals are out of scope for the current evaluation.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit one shot op
    For GDPval, we provide the full context of the task in the prompt, but in real life it often takes effort to figure out the full context of a task and understand what to work on.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit cost op
    Constructing and running our evaluation is expensive, particularly with industry expert graders. For this reason, we make an automated grader proxy available, but do not consider it a full substitute for industry expert graders.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit grader self op
    Our automated grader, based on GPT-5-high, shows lower correlation with human expert graders when assessing outputs from capable OpenAI models. This aligns with empirical evidence that models often favor their own responses
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit savings op
    We do not include consideration of the time taken to review a human professional deliverable, although this would commonly occur for tasks in GDPval (either self-review of the professional’s own work or review by a supervisor of a team member’s work). We also do not include the possibility that the human deliverable is also undesirable. One further limitation of this analysis is that it does not capture the cost of catastrophic mistakes, which can be disproportionately expensive in some domains.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit self report op
    Because times were self-reported, it is possible that experts under-estimated or over-estimated time taken
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit blinding op
    We aimed to keep comparisons as blind as possible, but model samples may still have been identifiable due to stylistic differences.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • limit under context op
    As a note: this experiment was run on an earlier version of the GDPval gold subset, and therefore the observed win rates do not match those in the main text of the paper.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • data access op
    We open-source the prompts and reference files in our 220-task gold subset. While human expert comparison is still our recommended method of grading, we make an experimental automated grader publicly available at evals.openai.com . Please note that the tasks in the open sourced set have been scrubbed of information that could be used to identify the expert who wrote the task.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • data access ungradable op
    In the open-source set we mark 12 out of 220 tasks as ungradable due to limitations of the automated grader.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • license op
    License: CC BY 4.0
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • lifecycle op
    Future GDPval iterations will incorporate greater breadth, realism, interactivity, and contextual nuance.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • version op
    arXiv:2510.04374v1 [cs.LG] 05 Oct 2025
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
  • held out 3p
    The held-out set of 1,100 tasks is only available internally in OpenAI, and external models cannot be submitted for grading on this set.
    https://en.wikipedia.org/wiki/AI_agent_benchmark · read 2026-10-29 · saved copy sha256 3be91030981967b0…
  • hf dataset op
    220 real-world knowledge tasks across 44 occupations. Each task consists of a text prompt and a set of supporting reference files.
    https://huggingface.co/datasets/openai/gdpval · read 2026-10-29 · saved copy sha256 1de40771dafd9ab6…
  • hf deliverables rubrics op
    deliverable_files listlengths 0 6 | deliverable_file_urls listlengths 0 6 | deliverable_file_hf_uris listlengths 0 6 | rubric_pretty stringlengths 1.15k 15.8k | rubric_json
    https://huggingface.co/datasets/openai/gdpval · read 2026-10-29 · saved copy sha256 1de40771dafd9ab6…
  • hf rubric author op
    "author_type": "human"
    https://huggingface.co/datasets/openai/gdpval · read 2026-10-29 · saved copy sha256 1de40771dafd9ab6…
  • leaderboard 3p
    We note that the official score reported in the official GDPval leaderboard by OpenAI reports a lower score of 31.9% (wins only) and 34.6% (wins and ties).
    https://ukgovernmentbeis.github.io/inspect_evals/evals/gdpval/index.html · read 2026-10-29 · saved copy sha256 db6622b4af380504…
  • leaderboard run 3p
    We ran our implementation for GPT-5 with thinking set to low, submitted it to OpenAI, and obtained the following results.
    https://ukgovernmentbeis.github.io/inspect_evals/evals/gdpval/index.html · read 2026-10-29 · saved copy sha256 db6622b4af380504…
  • leaderboard run score 3p
    Average Score: 47.3% ± 6.2%
    https://ukgovernmentbeis.github.io/inspect_evals/evals/gdpval/index.html · read 2026-10-29 · saved copy sha256 db6622b4af380504…
  • grading service 3p
    Note that OpenAI provides their automated grader model for free.
    https://ukgovernmentbeis.github.io/inspect_evals/evals/gdpval/index.html · read 2026-10-29 · saved copy sha256 db6622b4af380504…
  • grading service form 3p
    After uploading, the user should submit the OpenAI form to request grading.
    https://ukgovernmentbeis.github.io/inspect_evals/evals/gdpval/index.html · read 2026-10-29 · saved copy sha256 db6622b4af380504…
  • cdn pdf op
    Our paper introduces the first version of GDPval, a benchmark evaluating AI model performance on real-world economically valuable tasks.
    https://cdn.openai.com/pdf/d5eb7428-c4e9-4a33-bd86-86dd4bcf12ce/GDPval.pdf · read 2026-10-29 · saved copy sha256 37c7604bc9da68cd…
  • cdn pdf created op
    CreationDate: Thu Sep 25 17:03:48 2025 EDT
    https://cdn.openai.com/pdf/d5eb7428-c4e9-4a33-bd86-86dd4bcf12ce/GDPval.pdf · read 2026-10-29 · saved copy sha256 37c7604bc9da68cd…
  • builds on usage op
    We also build on production AI usage analyses (e.g., Tamkin et al. (2024) ; Chatterji et al. (2025) ; Appel et al. (2025) ) to cover areas where model adoption is still emerging.
    https://arxiv.org/html/2510.04374v1 · read 2026-10-29 · saved copy sha256 1d4c4f2494000318…
Negative tagswhat it is not5
  • NOT a measure of AI use in the economy: a benchmark of model capability on constructed tasks, offered by its authors as an alternative to lagging adoption indicators
  • NOT an employment outcome measure
  • NOT a test of interactive or under-specified work, except one experiment on an earlier gold subset
  • deliverable parity with experts is NOT a finding of job replacement in the paper
  • NOT independent of a provider: built and graded by OpenAI, with an automated grader based on an OpenAI model
Sourcesevery source read, with its date and hash7
Not yet checkedwhat remains to read7
  • OpenAI's own post (openai.com/index/gdpval/), which answered 403 and is unread
  • the paper's figures (charts are images; the per-model win rates of Figures 5, 6 and 10-13 are read only where the text or Table 2 gives them)
  • the expert compensation rate and the vendor partners, which the paper does not state
  • the gold subset's files themselves (the dataset page and its preview rows were read; the files were not opened)
  • OpenAI's grading service at evals.openai.com (described by Inspect Evals, third party; not opened)
  • OpenAI's leaderboard page (its figures are known here only as Inspect Evals reports them)
  • an internal difference in the paper: the full set is described as 'at least 30 tasks per occupation' and as '30 total tasks per occupation' (1,320 = 44 x 30); and 47.6% for the best model sits beside 'just over half the tasks' in the same paragraph

Cite this record

Lee Sharks, ed. "GDPval." Labor Vector meta-aggregator, laborvector.org, compiled 2026-10-29. https://www.laborvector.org/aggregators/openai-gdpval

Record https://www.laborvector.org/aggregators/openai-gdpval · entity https://www.laborvector.org/aggregators/openai-gdpval#entity · the census as data

@misc{lv-openaigdpval2026,
  editor = {Sharks, Lee},
  title = {GDPval},
  howpublished = {Labor Vector meta-aggregator, laborvector.org},
  year = {2026},
  note = {Compiled 2026-10-29; every field sourced and dated},
  url = {https://www.laborvector.org/aggregators/openai-gdpval}
}
Editor: Lee Sharks (ORCID 0009-0000-1599-0703). This page is the Labor Vector meta-aggregator's record of the resource above: its reading and fields are ours, and each field carries its source, its basis and the hash of the copy read. The resource's own terms and findings are kept as its own. Compiled 2026-10-29. CC BY 4.0.