No AI agent scores above 12 percent on SemiAnalysis' new real-world work benchmark

SemiAnalysis has introduced the GDPx benchmark to evaluate AI agents on complex knowledge tasks, such as chip design and quantitative trading. The results indicate that even the top-performing models, including Claude Opus 5.5 and Gemini 4 Carbon, currently fail to complete more than 12 percent of assigned work correctly. This low success rate underscores the significant limitations of current agentic systems in executing multi-step, real-world professional workflows.