GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains

SemiAnalysis has published an article introducing GDPx, an evaluation that, in the authors' description, measures frontier model capabilities across what the piece calls the most valuable domains. The article appears on the SemiAnalysis newsletter. It describes GDPx as assessing whether AI agents can autonomously perform the research workflows actually used inside frontier labs, rather than the general-purpose tasks that make up conventional benchmarks.

The workflows named in the article include identifying reward hacking, debugging failed training runs, improving data quality and improving inference algorithms. The authors argue that general-purpose benchmarks no longer capture where frontier progress is happening. They cite OpenAI's GDPval, released over a year ago, as an example: the article states that a benchmark of that kind, issued more than twelve months earlier, has been outpaced by progress at the frontier.

GDPx is presented as measuring capabilities across the domains the authors identify as the most valuable, with autonomous completion of lab research as the property under test. The article's introduction sets out the design and the rationale for the evaluation; it does not, in the passage provided, report any model's scores. The comparison drawn is with GDPval, attributed to OpenAI and described as having been released over a year before this piece. The piece is framed as an argument that evaluation should follow the work researchers actually do inside labs, rather than general-purpose task sets.