How close are frontier AI models to expert work quality?

The GDPval paper states that "the current best frontier models are approaching industry experts in deliverable quality," and that frontier performance "is improving roughly linearly over time." The benchmark covers most BLS work activities for 44 occupations in the top 9 GDP-contributing sectors, with 1,320 tasks in the full set and at least 30 per occupation. Each task is built from the work product of an industry professional with at least four years in the occupation (14 on average). The primary metric is "head-to-head human expert comparison," and each task received an average of five human reviews. The excerpt omits the win-rate tables behind the "approaching" claim, so that claim rests on the abstract's wording here.

The construction is economic on purpose. Sectors are ranked by GDP contribution, occupations are drawn from the highest-earning knowledge-work roles in them, and each task's dollar value is its average completion time multiplied by median hourly wages from OEWS data. The paper also treats the scores as a property of how models are run. For o3 and GPT-5, reasoning effort from low to high improved performance. A prompt asking GPT-5 to check deliverables, render layouts as images and avoid nonstandard Unicode raised human preference win rates by five percentage points and removed black-square artifacts that had affected over half of its PDFs. PowerPoint files with egregious formatting errors fell from 86% to 64%, which the paper partly credits to agents inspecting their own output with multimodal tools, a rise from 15% to 97%. The authors call these "easy performance gains" that point toward training or scaffolding agents to be more thorough.

The nearest note on this gap argues that the benchmark-to-GDP gap is an evaluation artifact: agents clear contests but not the long-horizon occupational workflows that pay, and the hardest ALE tier sits below 1% full pass. GDPval is also an occupational benchmark, and it reports the opposite picture. The difference looks mostly like scoring design. ALE is graded deterministically, while GDPval uses human expert preference on tasks that are fully specified up front. So this excerpt does not overturn the ALE note. It shows that the same kind of expert-built task can read as near-expert or near-zero depending on how it is graded and how much context is handed over. The paper's limitations section names the distortion that open-world evaluation targets: GDPval tasks "are precisely-specified and one-shot, not interactive," and in real work "it often takes effort to figure out the full context of a task." The abstract's cheaper-and-faster analysis is conditioned on "human oversight," which is where the accountability question from What makes accountable judgment scarce when AI cognition is cheap? returns. The excerpt does not show how that oversight is done or what it costs.

What the excerpt does not establish is the cost and speed numbers, the win rates behind "approaching," and any measure of how these deliverables are used. The scope limits are the paper's own: 44 occupations with about 30 tasks each, which it calls "a limited, initial cut," and no manual work, no tasks needing extensive tacit knowledge, personal data, proprietary software or communication between individuals. The tested models are the paper's own, and the grader is hosted at evals.openai.com. So the result is best read as a capability measurement under one-shot prompting, where prompt and scaffold changes move the score. It supports a claim about digital deliverables in these occupations. It does not, on its own, support a claim that AI can do the jobs those occupations contain.