Advanced AI Models Score Maximum 12 Percent on New Professional Task Benchmark
November 6, 2026 · 2 min read
State of the art AI models achieved a maximum score of 12.0 percent in an evaluation designed to replicate professional corporate tasks.
SemiAnalysis released a new AI agent benchmark called GDPx. The evaluation required the performance of 206 tasks across 11 professional fields, including semiconductor design, cybersecurity, biology, finance, law, software development, and AI self improvement. To create these tests, SemiAnalysis collaborated with 14 specialized firms to recreate actual work environments.
Claude Opus 5.5 from Anthropic ranked first with a score of 12.0 percent. Gemini 4 Carbon from Google followed with 10.9 percent, while GPT-6 Astra from OpenAI scored 9.6 percent and Muse Wonderful from Meta scored 9.0 percent. These figures are composite scores reflecting partial points and complete success across tasks, with each task given up to 8 hours and averaged over three executions.
SemiAnalysis developed GDPx to measure the ability to create economic value in actual industries rather than general intelligence. The benchmark focuses on high value areas such as finance and semiconductor design because the company believes initial revenue from AI will be concentrated in professional industries with high automation effects. The results will be used to predict the revenues of OpenAI and Anthropic for 2027 to 2028.
The benchmark identifies limitations in previous evaluations like GDPval, which SemiAnalysis claims focused too much on writing reports based on provided instructions. In contrast, GDPx requires AI to find necessary information from vast data, compare conflicting content, and judge conditions not explicitly instructed. For a finance task, the benchmark created a virtual environment with 4,000 anonymized files based on actual corporate acquisitions, requiring the AI to find and analyze financial statements and contracts.
Many models failed to complete complex tasks. In one finance task analyzing corporate profitability, no model achieved complete success across 25 executions, with the best performance passing only 8 of 11 verification items. In the legal field, four models failed to meet completion standards across 20 attempts to edit investment contract documents, missing loopholes or adding conflicting conditions.
Performance varied by field. Opus 5.5 scored highest in hardware design, semiconductor design, and quant trading. Gemini 4 Carbon led in software development and biology, while GPT-6 Astra ranked first in materials science and AI self improvement.
Limitations of the evaluation include the use of different agent execution environments, such as Claude Code and Codex, for different models. Additionally, safety guards restricted tasks in some biology and cybersecurity tests, and some scoring criteria remain undisclosed.