What happened
OpenAI announced GDPval on September 25, 2025, describing it as an evaluation intended to track how its models and others perform on economically valuable, real-world work. The first version covers 44 occupations selected from nine industries contributing over 5% each to U.S. GDP. OpenAI says it selected five occupations per industry based on wage and employment data, then focused on those where at least 60% of component tasks were classified as knowledge work rather than physical work or manual labor.
The full set contains 1,320 specialized tasks, including a 220-task gold open-source set. Each task was created and vetted by experienced professionals whose average experience exceeded 14 years. Examples of the work products underlying tasks include legal briefs, engineering blueprints, customer-support conversations and nursing care plans. The expected outputs can include documents, slides, diagrams, spreadsheets and multimedia, and tasks come with reference files and context rather than being only a standalone prompt.
The benchmark is designed to broaden evaluation beyond academic tests, coding challenges and narrow economic-task benchmarks. OpenAI places GDPval in a progression that includes evaluations such as SWE-Bench, MLE-Bench, Paper-Bench and SWE-Lancer. Its stated distinction is breadth across occupations and industries, combined with tasks built around real or plausibly constructed work products. The company presents this as a way to ground claims about AI’s workplace potential in observed performance rather than speculation.
GDPval’s occupation selection is itself a scoped design choice, rather than a census of work. OpenAI began with industries that each contribute more than 5% to U.S. GDP, using Federal Reserve Bank of St. Louis data, and then used May 2024 U.S. Bureau of Labor Statistics wage and employment data in its selection process. It used O*NET task data to distinguish knowledge work from work requiring action in the physical world. That makes the benchmark most directly a measure of a particular slice of U.S. professional work, not a general measure of all jobs or all forms of economic output.
The source links to a research paper and an evaluation site, but the supplied material does not specify a price or commercial access terms. It says the gold subset is open-sourced and that the full set comprises 1,320 tasks. The announcement describes the first version; it says future versions will expand coverage and move toward more interactive workflows and richer context. No new version or later release is described here.
What the archive adds
A Telegram post dated 2025-09-27 already described GDPval’s 44 professions and 1,320 tasks. New Runtime also covered SemiAnalysis’s GDPx benchmark, which spans 11 domains and 206 tasks. This is therefore a restatement of a known benchmark, with the useful distinction that GDPval is an occupational work-product evaluation while GDPx is an agentic benchmark.
Trend signal
This story fits the pattern that economically meaningful agent progress is measured by completed work rather than model calls: Agent economics moves to completed work. GDPval makes work products and occupational tasks the unit of evaluation. However, its one-shot setup does not yet establish how an agent performs across a full task-completion loop, including clarification, checking, recovery from mistakes, and delivery to a responsible human.
Why now
OpenAI frames GDPval as part of a progression from academic evaluations toward more applied and economically relevant tasks. The announcement argues that forecasts about AI’s social and workplace effects need evidence about what models can already do. Its proposed next step is not simply more difficult questions, but evaluations based on artifacts and context that resemble professional work. The source also says future versions are intended to add interactive workflows and richer context.
Connections to other events
The archive’s GDPx coverage described an agentic benchmark across 11 high-value domains and 206 tasks, where the reported leading score was 12%. Both efforts move evaluation toward applied work, but they make different units of measurement central: GDPval emphasizes occupational deliverables across a broad set of professions, while GDPx is described as agentic and domain-focused. Neither comparison alone resolves whether performance transfers to a team’s actual operating environment.
Why it matters
For a team deciding whether to use AI in professional workflows, a benchmark built around realistic artifacts is a better starting point than a model’s score on an abstract exam. GDPval’s breadth can help frame the question of where models may be useful across knowledge work, and its file-based tasks resemble the materials teams actually handle. It also offers a more concrete vocabulary for internal evaluation: task inputs, context files, a required artifact, and a reviewable result.
The benchmark is not a deployment decision by itself. A high score on a task set does not show that a system can reliably handle a team’s permissions, proprietary context, multi-step handoffs or review requirements. Builders can use the benchmark as a prompt to define representative internal tasks and acceptance criteria, then evaluate systems against those criteria in their own environment. The one-shot limitation is especially important: many jobs involve clarification, revision and accountable sign-off, not one answer to one prompt.
Teams should also separate artifact quality from operational fitness. A polished document or spreadsheet may still be unsuitable if its inputs were incomplete, its claims cannot be checked, or it was produced without the controls a workflow requires. GDPval is useful for narrowing the gap between test prompts and professional outputs; it does not remove the need for local evaluation, human review, and clear responsibility for use of the result.
What happens next
- A later GDPval iteration will add more interactive or context-rich workflows. (12 months, 60%). Watch for: an OpenAI announcement describing multi-draft, multi-turn, or richer-context tasks.
- Teams will use artifact-based internal evaluations alongside public benchmark scores. (6 months, 65%). Watch for: evaluation suites built around company documents, files, and explicit acceptance criteria.
Limits and uncertainty
OpenAI explicitly says GDPval is an early step and does not capture the full nuance of economic tasks. Its one-shot design omits situations where a model must build context or improve through multiple drafts. The supplied material does not report a model ranking, scores, error rates, evaluation methodology details, or evidence that benchmark performance translates into productivity gains, lower costs or labor-market effects.
It also does not establish how representative the selected U.S. occupations are of work in other countries or of physical and service work. The selection process prioritizes occupations judged predominantly to involve knowledge work, and the announcement does not show whether individual task outcomes vary materially by occupation, artifact type, or model configuration. Plans for broader and more interactive future versions are plans, not reported outcomes.
Skill to build now
Create a small internal artifact evaluation this week. Choose three recurring tasks that each produce a tangible output, such as a customer response, a project brief, or a spreadsheet analysis. For each, preserve the normal context a colleague would receive, define what an acceptable output must contain, and have a domain reviewer compare AI output with the expected standard. Record revision requests and factual failures separately from style preferences. That exercise tests the gap GDPval identifies: producing a plausible artifact is different from completing work to a team’s standard.
Reader takeaway
GDPval is useful because it evaluates models on realistic professional deliverables across many occupations, not just exam-style questions. Treat it as evidence about performance on a defined, one-shot task set—not proof that AI can independently complete a job or improve workplace productivity. For builders, the strongest next step is to test systems on representative tasks with real context and explicit review criteria.
Open question
Will future GDPval versions test multi-turn workflows with revisions and richer context, and how will those changes affect model comparisons?