SemiAnalysis has released GDPx, a new benchmark that tests frontier AI models on agentic tasks across 11 economically valuable domains, including chip design, biology and finance. The headline result is a humbling one: the best model on the v1.0 leaderboard, Anthropic's Claude Opus 5.5, succeeded on just 12.0% of tasks.

Behind it sit Google's Gemini 4 Carbon at 10.9% and OpenAI's GPT-6 Astra at 9.6% — all far below full task completion. The spread is not uniform: Gemini 4 Carbon led in biology and software engineering, while GPT-6 Astra led in recursive self-improvement and materials science. SemiAnalysis picked the domains by where it believes frontier intelligence will earn the highest return on investment, rather than by GDP share, and built the tasks with 14 partner startups. The benchmark excludes the government and insurance sectors.

The takeaway for anyone buying AI tools for professional work is blunt: on complex, multi-step work, today's frontier models still need substantial human oversight. GDPx is designed to be updated every three to six months, so it is worth watching as a tracker of where that gap is actually closing — and where vendors' marketing runs ahead of measured task success.

Sources