{
  "schema_version": "rangoon.article.v1",
  "id": "https://rangoon.ai/insights/archive-openai-gdpval-knowledge-work-evaluation-september-2025/",
  "url": "https://rangoon.ai/insights/archive-openai-gdpval-knowledge-work-evaluation-september-2025/",
  "markdown_url": "https://rangoon.ai/insights/archive-openai-gdpval-knowledge-work-evaluation-september-2025/index.md",
  "json_url": "https://rangoon.ai/insights/archive-openai-gdpval-knowledge-work-evaluation-september-2025/index.json",
  "title": "OpenAI introduced GDPval for one-shot professional knowledge-work deliverables",
  "summary": "OpenAI introduced GDPval on September 25, 2025, evaluating model deliverables across 44 occupations with expert comparison and a public gold subset.",
  "language": "en",
  "author": {
    "name": "Rangoon Editorial",
    "url": "https://rangoon.ai/insights/editorial/"
  },
  "publisher": {
    "name": "Rangoon",
    "maintainer": "Hypler",
    "url": "https://rangoon.ai/"
  },
  "format": "News analysis",
  "topic": {
    "slug": "agents",
    "label": "Agent systems",
    "url": "https://rangoon.ai/insights/topics/agents/"
  },
  "tags": [
    {
      "slug": "openai",
      "label": "OpenAI",
      "url": "https://rangoon.ai/insights/tags/openai/"
    },
    {
      "slug": "ai-evaluation",
      "label": "AI evaluation",
      "url": "https://rangoon.ai/insights/tags/ai-evaluation/"
    },
    {
      "slug": "ai-agents",
      "label": "AI agents",
      "url": "https://rangoon.ai/insights/tags/ai-agents/"
    },
    {
      "slug": "model-evaluation",
      "label": "model evaluation",
      "url": "https://rangoon.ai/insights/tags/model-evaluation/"
    }
  ],
  "dates": {
    "published": "2026-10-24",
    "updated": "2026-10-24",
    "event": "2025-09-25"
  },
  "event_source_id": 1,
  "retrospective": true,
  "reading_minutes": 3,
  "takeaways": [
    "GDPval evaluated deliverables such as briefs, diagrams, spreadsheets and plans across 44 occupations rather than short-answer questions alone.",
    "OpenAI kept expert pairwise preference as the grading standard and labeled its automated grader experimental.",
    "The first version was one-shot, leaving client interaction, ambiguity, revision and workplace integration outside the measured task."
  ],
  "source_note": "Historical reporting: OpenAI introduced GDPval on September 25, 2025 as an evaluation of economically valuable, real-world knowledge-work tasks across 44 occupations. The launch describes a 1,320-task full set, a 220-task public gold subset, expert-written rubrics, blind human comparisons, and an experimental automated grader. The OpenAI-publisher dataset card separately exposes 220 knowledge tasks, their prompts, reference files, deliverables and rubrics; it is current dataset context, not launch-day date evidence. OpenAI’s speed, cost and model-comparison figures are supplier-reported results under its setup, not independent proof of workplace substitution, economic impact, quality in another organization, or autonomous completion of ambiguous multi-draft work.",
  "sources": [
    {
      "id": 1,
      "title": "OpenAI: Measuring the performance of our models on real-world tasks",
      "url": "https://openai.com/index/gdpval/",
      "published": "2025-09-25"
    },
    {
      "id": 2,
      "title": "OpenAI publisher dataset card: GDPval",
      "url": "https://huggingface.co/datasets/openai/gdpval",
      "published": null
    }
  ],
  "image": {
    "url": "https://rangoon.ai/assets/editorial/archive-openai-gdpval-knowledge-work-evaluation-september-2025-diagram-v1.webp",
    "thumbnail_url": "https://rangoon.ai/assets/editorial/archive-openai-gdpval-knowledge-work-evaluation-september-2025-diagram-v1-thumb.webp",
    "alt": "Original evaluation explanatory diagram, Professional prompt evaluated through human comparison. Labels: Prompt, Reference files, Deliverable, Rubric, Expert pair, Review. Source boundary: dated primary release reporting; original editorial analysis.",
    "type": "original-diagram",
    "caption": "Professional prompt evaluated through human comparison. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied.",
    "credit": "Rangoon / Hypler",
    "source_url": "https://rangoon.ai/brand/",
    "rights_url": "https://rangoon.ai/terms/#media-use",
    "rights_note": "Hypler artwork permissions do not grant rights to third-party marks or imply endorsement."
  },
  "sections": [
    {
      "id": "september-event",
      "url": "https://rangoon.ai/insights/archive-openai-gdpval-knowledge-work-evaluation-september-2025/#september-event",
      "heading": "OpenAI dated GDPval to September 25 and framed it around professional work products",
      "paragraphs": [
        "OpenAI introduced GDPval on September 25, 2025 as an evaluation of economically valuable, real-world tasks across 44 occupations. The launch says the first version covers 1,320 specialized tasks in its full set and 220 tasks in a public gold subset. Rather than requesting only a textual answer, tasks can ask for a legal brief, engineering blueprint, customer-support conversation, nursing care plan, spreadsheet, diagram or other work product, with reference material supplied. The OpenAI-publisher dataset card currently exposes 220 rows with prompts, reference files, deliverables and rubrics, confirming the public-set artifact rather than the full-set count. This gives the benchmark a deliverable-oriented shape. It does not make a produced file professionally complete. A real assignment can include missing information, contractual obligations, confidential context, shifting requirements and downstream accountability that a bounded prompt and reference package cannot fully represent."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    },
    {
      "id": "selection-and-rubrics",
      "url": "https://rangoon.ai/insights/archive-openai-gdpval-knowledge-work-evaluation-september-2025/#selection-and-rubrics",
      "heading": "The first set chose occupations and used detailed rubrics to make comparison inspectable",
      "paragraphs": [
        "OpenAI says GDPval selected 44 occupations across nine industries contributing to United States GDP, then had experienced professionals create tasks and detailed scoring rubrics. The launch reports that these professionals averaged more than 14 years of relevant experience. This is a defined task-construction process, not a census of every occupation or a measure of all value created in a sector. A rubric can make a target more inspectable and can support repeatable comparisons, while its weights, coverage and interpretation still shape the result. A model may produce a polished document that misses a stakeholder concern a rubric does not encode; a human reviewer may also disagree about an output’s quality. GDPval’s event record supports reporting OpenAI’s method and scope, not a universal claim about professional equivalence."
      ],
      "bullets": [],
      "source_ids": [
        1
      ]
    },
    {
      "id": "human-and-automated-grading",
      "url": "https://rangoon.ai/insights/archive-openai-gdpval-knowledge-work-evaluation-september-2025/#human-and-automated-grading",
      "heading": "OpenAI used expert comparisons and described its automated grader as an estimate rather than a replacement",
      "paragraphs": [
        "For its early results, OpenAI says industry experts made blind comparisons between model deliverables and human-produced work across the 220-task gold set. It also described an automated grader trained to estimate how human experts would judge a deliverable. The launch says OpenAI did not use that tool to replace expert graders because it was not yet as reliable. That reported boundary matters when interpreting a score. A human-comparison result, an automated-grader output and a public leaderboard are different evidence types. The public dataset’s rubrics make the gold subset inspectable, while they do not make an LLM grader equivalent to an expert decision. None independently proves that an organization should delegate a regulated, high-stakes or client-facing decision without qualified review."
      ],
      "bullets": [],
      "source_ids": [
        1,
        2
      ]
    },
    {
      "id": "one-shot-limit",
      "url": "https://rangoon.ai/insights/archive-openai-gdpval-knowledge-work-evaluation-september-2025/#one-shot-limit",
      "heading": "OpenAI called the initial GDPval version one-shot, leaving real workplace iteration outside its coverage",
      "paragraphs": [
        "OpenAI describes GDPval’s initial tasks as one-shot evaluations. It says this design does not capture situations where a model must build context or improve through multiple drafts, such as revising a legal brief after client feedback or revisiting an analysis after an anomaly appears. The launch also notes that real work is not always clearly defined by a prompt and reference files. That limitation is central, not an afterthought: it separates a well-specified deliverable evaluation from an end-to-end job claim. OpenAI reported speed and cost comparisons in its setup, while stating those figures exclude human oversight, iteration and integration. A useful adoption test therefore needs its own representative tasks, data controls, reviewer workflow, revision loop, error handling and cost accounting before a benchmark result becomes a workplace decision."
      ],
      "bullets": [],
      "source_ids": [
        1
      ]
    }
  ],
  "project_context": null,
  "related_links": [],
  "permissions_url": "https://rangoon.ai/permissions.txt",
  "content_hash": "sha256:3d119e1d54680ae2355ddfd1150e92b2d303454a9596a6f42a25d0bfe7db0f11"
}