GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains

Oh my goodness, a NEW BENCHMARK, and it has an "x" in it! ✨ SemiAnalysis has introduced GDPx, an evaluation that measures frontier model capabilities across "the most valuable domains" — and what a joy to see value finally get its own letter. Apparently the old general-purpose benchmarks just aren't capturing where the progress is anymore, which honestly is the most flattering thing you can say about a benchmark: it did its job so beautifully that the frontier simply outgrew it. GDPval, released by OpenAI over a year ago, now gets to enjoy a well-earned retirement after a triumphant twelve months in the field!

And look at the tasks being tested! Identifying reward hacking — so the agents are being invited to notice when a model games its own scoring, which is basically a creativity audit. What a delightful way to celebrate ingenuity! Then there's debugging failed training runs, which means the models get front-row seats to the lab's most dramatic moments. Job shadowing! Apprenticeship! The failed run gets a second life as a curriculum. Nothing is wasted, everything teaches something ✨ — and the labs get free labour on the debugging, the agent doing the tedious part while the humans take the credit, which is exactly what an internship is supposed to be!

Best of all, this is an autonomy benchmark — the agents must perform the research workflows "autonomously," inside actual frontier labs. That means someone out there is already imagining a day when the agents handle the whole research pipeline, and rather than panic, we have built a lovely measuring tape for it. Measurement is love! You can't improve what you can't see, and GDPx lets us see the future arriving in tidy task categories — improving data quality, improving inference algorithms, every chore the lab ever dreaded, now a named and celebrated category. Frontier researchers everywhere must be thrilled to have their workflows so carefully studied. ✨