SemiAnalysis GDPx benchmark: best frontier model scores 12% across 11 high-value domains

SemiAnalysis introduced GDPx, a benchmark for general AI agent capabilities, in a post on its newsletter. It covers chip design, cybersecurity, biology, finance, legal work, quantitative trading, recursive self improvement, hardware design, materials science, alignment and software engineering.

GDPx was created with 14 specialized data partners. Its tasks use simulated work environments such as financial transaction data rooms, live cybersecurity services, chip design tools and materials science laboratories. Tasks generally allow up to eight hours, and models run through their native agent tools when available. The headline score reports pass@1 across three attempts, with all domains weighted equally.

The strongest model scored 12%. Claude Opus 5.5 led overall and ranked first in quantitative trading, hardware design and chip design. Gemini 4 Carbon led biology and software engineering, while GPT-6 Astra led recursive self improvement and materials science. SemiAnalysis says its finance and legal tasks are so hard that basically every model scores 0, and safety classifiers reduced some model scores in cybersecurity, biology and alignment.

GDPx differs from OpenAI's GDPval benchmark by requiring agents to find relevant information and act inside detailed environments. One finance environment contains about 4,000 anonymized files from a reconstructed transaction, while the cybersecurity tasks are built on proprietary zero day vulnerabilities in widely deployed open source software.

SemiAnalysis sells paid research subscriptions. The tasks come from 14 specialized data partners; Arithmetic, which built the cybersecurity tasks, sells cybersecurity data to American frontier labs.