New AI test of expert-level work finds even the best model passes only about 12% of tasks
Research firm SemiAnalysis built GDPx from real tasks in fields such as chip design, biology and finance. Models scored near zero in finance and law under its strict pass-or-fail grading.
The best AI model passed only about 12% of tasks on a new test built from real expert work, according to SemiAnalysis, a chip and AI research firm that released the benchmark, called GDPx, on Nov. 2.
A benchmark is a standard test that lets people compare AI models. Many popular ones use short quiz-style questions that top models now ace. GDPx instead gives models long projects in 11 fields, including chip design, cybersecurity, biology, finance, law, software engineering and materials science, written with 14 specialized data companies, SemiAnalysis said.
Each model works for up to eight hours inside its own company's coding tool, such as Anthropic's Claude Code or OpenAI's Codex, and a task counts only if the work fully passes. Each field counts equally in the overall score.
SemiAnalysis said Anthropic's Opus 5.5 was the best all-around model, Google's Gemini 4 Carbon led in biology and OpenAI's GPT-6 Astra led in materials science and in tasks where AI improves AI. Under the strict grading, models scored close to zero in finance and law.
The firm said GDPx responds to OpenAI's 2025 GDPval test, which it said lacked realistic work environments, and plans new versions every three to six months. Results like these suggest AI tools still need close human checking on complex professional work.
Source: SemiAnalysis, "GDPx: Measuring Frontier Model Capabilities," Nov. 2, 2026; GDPx results site