On Thursday, OpenAI introduced a new benchmark designed to evaluate how its AI models stack up against human professionals across numerous sectors and job roles. This assessment, called GDPval, represents an initial step toward gauging how close OpenAI’s technology is to surpassing humans in economically significant tasks—an essential aspect of the company’s original goal to achieve artificial general intelligence (AGI).
According to OpenAI, its GPT-5 model and Anthropic’s Claude Opus 4.1 “are already nearing the level of work produced by top industry professionals.”
However, this doesn’t mean OpenAI’s models will immediately take over human jobs. While some business leaders predict AI will replace human workers within a few years, OpenAI concedes that GDPval currently only evaluates a narrow set of tasks that people perform in their actual jobs. Still, it serves as one of the latest metrics the company uses to track AI’s advancement toward this goal.
GDPval focuses on nine major industries that make up the largest share of the United States’ gross domestic product, covering areas like healthcare, finance, manufacturing, and government. The benchmark measures AI performance across 44 different professions within these fields, including roles such as software developers, nurses, and journalists.
For the initial version, GDPval-v0, OpenAI enlisted seasoned professionals to review and compare reports generated by AI with those written by their peers, selecting which they found superior. For instance, one scenario had investment bankers create a competitive analysis for the last-mile delivery sector and compare it to AI-generated reports. OpenAI then calculated the average “win rate” of the AI model against human-created reports across all 44 professions.
With GPT-5-high, an enhanced version of GPT-5 with greater computational resources, OpenAI reports that the model matched or exceeded expert-level work in 40.6% of cases.
Anthropic’s Claude Opus 4.1 was also evaluated, with results showing it performed as well as or better than industry experts in 49% of tasks. OpenAI attributes Claude’s high score in part to its ability to produce visually appealing graphics, rather than just raw performance.

It’s important to recognize that most professionals do far more than just submit research reports, which is the sole focus of GDPval-v0. OpenAI acknowledges this limitation and intends to develop more comprehensive benchmarks in the future that can better reflect a wider range of industries and more interactive job functions.
Even so, the company considers the progress shown by GDPval to be significant.
Speaking with TechCrunch, OpenAI’s chief economist Dr. Aaron Chatterji noted that GDPval’s findings indicate people in these roles can now leverage AI to focus on more meaningful aspects of their work.
“[Since] the model is becoming proficient at some of these tasks,” Chatterji explains, “workers in these positions can increasingly rely on the model to handle certain responsibilities, freeing them up to pursue higher-value activities.”
Tejal Patwardhan, who leads evaluations at OpenAI, told TechCrunch she is optimistic about the rapid progress seen with GDPval. For example, OpenAI’s GPT-4o model, which launched about 15 months ago, achieved only a 13.7% win or tie rate against humans. Now, GPT-5’s results are nearly three times higher, a trend Patwardhan expects will continue.
The tech industry uses a variety of benchmarks to gauge the advancement of AI models and determine if a model is truly cutting-edge. Popular examples include AIME 2025, which tests competitive math skills, and GPQA Diamond, which assesses knowledge of PhD-level science. Yet, many AI models are reaching their limits on these tests, and researchers are calling for new benchmarks that better evaluate real-world capabilities.
Benchmarks like GDPval may play a growing role in this discussion, as OpenAI argues its models can benefit a broad spectrum of industries. Still, a more thorough version of the test may be necessary before OpenAI can definitively claim its AI surpasses human performance.