No model scored above 12% in the GDPx benchmark for AI agents

Analytics firm SemiAnalysis has introduced a new benchmark, GDPx, designed to evaluate the capabilities of AI agents in performing complex professional tasks. The testing covers a wide range of disciplines, including microchip design and quantitative trading in financial markets.

The study's results showed that current flagship models are failing to meet the set tasks: none of them managed to cross the 12% success rate threshold. Leading the rankings was Claude Opus 5.5, which showed a minimal advantage over Gemini 4 Carbon.

The GDPx methodology has faced criticism from experts. The main concerns relate to the limited number of tasks for each subject area, as well as uneven testing conditions, which may distort the real picture of neural network capabilities in workflows.

For developers and companies implementing autonomous agents, this data serves as an indicator that current systems are not yet ready to fully replace specialists in highly specialized technical fields. The low scores highlight the gap between the models' ability to generate text and their skill in solving multi-stage engineering tasks.