SemiAnalysis launches GDPx, an agentic capability benchmark measuring 11 domains, including chip design, cybersecurity, and RSI; Opus 5.5 tops with a 12% score

The best AI models score just 12% on this new benchmark, meaning current agents are far from human-level general intelligence for complex tasks.

SemiAnalysis has launched GDPx, a new benchmark designed to measure the general agentic capabilities of frontier AI models across 11 key domains. These domains include chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement (RSI), hardware design, materials science, alignment, and software engineering, deliberately excluding sectors like government and real estate. The benchmark aims to reflect the areas where AI is expected to generate the highest return on investment. In its initial release, even the top-performing model, Opus 5.5, achieved only a 12% score, indicating the nascent stage of general agentic AI. The benchmark also noted that safety classifiers can influence overall scores, and when these are excluded for specific domains, rankings can shift. Google's Gemini 4 Carbon shows significant improvement, positioning it competitively with leading models like Opus and OpenAI's GPT-6 Astra.

The GDPx benchmark signifies a shift towards evaluating AI agents in more realistic, domain-specific environments with rich context, moving beyond the limitations of older benchmarks like OpenAI's GDPval. The involvement of 14 specialized data partners highlights the difficulty and necessity of deep domain expertise in creating meaningful AI evaluations, a challenge that single entities often struggle to meet.

This development suggests a market that is rapidly professionalizing AI evaluation, with a focus on practical business impact rather than theoretical general knowledge. The current low scores across all models indicate that while agentic AI is advancing, achieving robust performance in high-ROI sectors still requires substantial further development and specialized tooling.