Research firm SemiAnalysis on Monday published GDPx, a benchmark it said is meant to measure general agentic capabilities of frontier AI models across 11 domains, including chip design, cybersecurity and recursive self-improvement¹.

On the v1.0 leaderboard, Anthropic's Claude Opus 5.5 posted the highest overall score at 12.0%, SemiAnalysis said, followed by Google's Gemini 4 Carbon at 10.9% and OpenAI's GPT-6 Astra at 9.6%¹.

The public GDPx site lists 11 domains, 14 data partners and 206 tasks, with each domain weighted equally and each partner within a domain also weighted equally¹.

SemiAnalysis said it chose domains it expects will absorb high-return uses of frontier models, rather than weighting tasks by share of GDP or occupational taxonomies used in older knowledge-work tests such as OpenAI's GDPval¹.

The firm said Opus 5.5 was the strongest all-around model and led in quant trading, hardware design and chip design, while Gemini 4 Carbon led in biology and software engineering and GPT-6 Astra led in recursive self-improvement and materials science¹.

Finance and legal tasks were so difficult that essentially every model scored zero, SemiAnalysis said¹.

The authors said some overall scores, including for Opus 5.5 and GPT-6 Astra, are held down by safety classifiers, particularly in cybersecurity, biology and alignment, and that Muse Wonder Full was the strongest cyber model in their tests¹.

Chip design is the only domain with three task partners — Aria, Arcverex and Ergodex — which SemiAnalysis said it overweighted because it expects AI spend on chip design to rise in 2027¹.

The firm said headline scores are pass@1 across three attempts per task, using a mix of binary and partial grading depending on the domain¹.

Sources

1. SemiAnalysisGDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains2:06 AM

Read the original at SemiAnalysis ↗