Photo: Tima Miroshnichenko / Pexels
The research firm's new test targets 11 high-value domains where frontier AI is expected to earn its first trillion dollars
SemiAnalysis has released a new AI benchmark called GDPx. The top-scoring model on its first leaderboard managed a 12.0% overall success rate.
The research firm unveiled GDPx on November 2, 2026. It describes the benchmark as an attempt to build the world’s best yardstick for general agentic capabilities, meaning AI systems that can take multi-step actions toward a goal rather than simply answer questions.
What GDPx actually measures #
GDPx evaluates frontier models across 11 domains. Among them are chip design, biology, and finance.
The selection is deliberate. SemiAnalysis argues that the first trillion dollars of AI revenue will cluster in sectors that can put frontier intelligence to work at high return on investment, rather than spreading evenly across the economy.
That thesis drives the whole design. Instead of weighting tasks by each industry’s share of GDP, or mapping them to the SOC/O*NET occupational taxonomy that classifies jobs across the workforce, the firm picked the areas where it believes frontier models will matter most.
The flip side is that some sectors were left out on purpose. Government and insurance are excluded from GDPx, as SemiAnalysis views them as offering lower ROI potential for AI applications.
To build the actual tasks, SemiAnalysis worked with 14 specialized startups that provide data and reinforcement learning services. Each partner helped generate challenges within specific domains of the framework.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The leaderboard, and why the numbers are so low #
Anthropic’s Claude Opus 5.5 took the top overall spot on the GDPx v1.0 leaderboard with a score of 12.0%.
Google’s Gemini 4 Carbon came second with 10.9. OpenAI’s GPT-6 Astra followed at 9.6. The overall ranking hides a more interesting story underneath. No single model dominated every category.
Gemini 4 Carbon led in both biology and software engineering. GPT-6 Astra took first place in recursive self-improvement and materials science.
According to the research findings, no current model comprehensively solves tasks across all of the evaluated domains. The leader in one specialty can look mediocre in another.
A benchmark built to change #
SemiAnalysis has said GDPx will not be static. The firm plans to update the benchmark every 3 to 6 months.
Those updates are expected to fold in feedback and bring on additional partners.
What this means for the AI race #
The headline ranking favors Anthropic. The domain-level results show that Google and OpenAI each hold clear advantages in specific technical areas, which complicates any simple claim about which lab is ahead.
The research findings suggest that top models clearing roughly one in ten tasks indicates that capabilities are advancing, but significant barriers remain before any model reaches broad proficiency in high-stakes professional work.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.