The best AI agent finishes 12% of real work tasks. Is that good or bad?
Everyone keeps saying AI is about to do all our jobs. So here is a number to hold next to that. SemiAnalysis just launched GDPx, a benchmark of 206 long tasks in 11 areas, from chip design and cybersecurity to biology, finance and legal work. Each task can run for up to eight hours.
The best score on the board? Claude Opus 5.5 at 12.0%. Gemini 4 Carbon is on 10.9%, GPT-6 Astra on 9.6%. Further down, an open-weight model sits at 2.7%. And in finance and legal, the leading agent scores zero.
So is AI overhyped?
Hold on. This benchmark was built to be brutal. It is meant to show how far the ceiling is, not to say AI is useless. A low number on a very hard test tells you where the frontier really is, not that nothing works.
And here is the twist nobody expects: the safety filters are costing points. Opus 5.5 and GPT-6 Astra sometimes refuse work in cybersecurity, biology and alignment. Take those three areas out and Astra edges ahead by a tenth of a point.
That makes the leaderboard a lot less tidy than the headline. There is also no single winner. Opus leads in quant trading, hardware design and chip design, Gemini does best in biology, and Astra wins in materials science and recursive self-improvement.
The honest takeaway: the models are very good at pieces of work, and still bad at carrying a whole long job from start to finish. If you are choosing an agent for a business, the right question is not "which is the best", it is "best at what".