SemiAnalysis launched GDPx, a benchmark for frontier model capabilities. It covers chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment, and software engineering. The best model scores just 12% overall. Safety classifiers on some models lower scores in cyber, biology, and alignment. Opus 5.5 leads in quant trading, hardware design, and chip design. Gemini 4 Carbon leads in biology and software engineering. GPT-6 Astra leads in recursive self-improvement and materials science. Finance and legal tasks are so hard that most models score zero. GDPx was built with 14 domain-specific partners. It replaces GDPval, which SemiAnalysis argues is outdated for agent evaluation. GDPx tasks include realistic, interactive environments. GDPval tasks are non-interactive and lack depth.
What it unlocks
Evaluating frontier models on realistic, domain-specific agentic tasks across high-ROI sectors.
- 12% best model score
- 14 specialized data partners
- $5k+ per high-quality task
Sources