SemiAnalysis launches GDPx benchmark, with top model scoring 12%

SemiAnalysis has launched GDPx, an agentic benchmark covering eleven economically significant domains, including chip design, cybersecurity, biology, finance, legal and software engineering. The best-performing model achieved 12% overall, while every tested model scored 0% in finance and legal under binary grading.

Opus 5.5 ranked best overall and led quant trading, hardware design and chip design. Gemini 4 Carbon led biology and software engineering, while GPT-6 Astra topped recursive self-improvement and materials science. Safety-classifier refusals may have depressed the overall results for Opus 5.5 and GPT-6 Astra, particularly in cybersecurity, biology and alignment. Muse Wonder Full, which refused fewer tasks, ranked highest in cybersecurity.

GDPx uses interactive environments rather than the precisely specified, non-interactive tasks used by OpenAI’s GDPval. Its finance environment, developed with Living Context, places models in a deal room containing roughly 2,000 files from a real private equity transaction representing over $100 billion in enterprise value. Third Moment contributed the quant trading component. The results highlight substantial limits in completing demanding professional work, although refusals complicate comparisons of underlying capabilities.