AI · November 6, 2026
New GDPx benchmark reveals top AI models complete only 12% of complex professional tasks
SemiAnalysis released a new evaluation framework called GDPx on November 2, 2026, which tests artificial intelligence agents on realistic professional work. The results showed that the highest-performing model achieved a score of 12.0 percent, indicating a significant gap between current AI capabilities and the demands of high-value industrial tasks. The benchmark was designed to measure the ability of AI to create economic value in specific sectors rather than general intelligence.
The evaluation covered 206 tasks across 11 specialized fields, including semiconductor design, cybersecurity, biology, finance, law, software development, and AI self-improvement. SemiAnalysis collaborated with 14 professional firms to recreate actual work environments for these tests. Each task allowed a maximum of eight hours for completion, and the final score was an average of three separate runs. The scoring system accounted for both complete success and partial credit, meaning the 12.0 percent figure does not imply that 12 out of 100 tasks were fully solved.
Anthropic's Claude Opus 5.5 led the rankings with a score of 12.0 percent. Google's Gemini 4 Carbon followed with 10.9 percent, OpenAI's GPT-6 Astra scored 9.6 percent, and Meta's Muse Wonder Full achieved 9.0 percent. The developers noted that high overall performance did not guarantee success in every specific domain. For instance, Claude Opus 5.5 excelled in semiconductor and hardware design, while Gemini 4 Carbon performed best in biology and software development. GPT-6 Astra led in AI self-improvement and materials science.
SemiAnalysis criticized existing benchmarks for failing to capture the complexity of expert judgment. They argued that previous tests, such as OpenAI's GDPval, often provided pre-organized data and clear instructions, which does not reflect how professionals work. Real experts must search through vast amounts of unstructured information, resolve conflicting data, and make judgments without explicit guidance. The new GDPx benchmark addresses this by requiring AI agents to find and analyze raw documents, such as financial statements and contracts, within a simulated environment.
In one financial task based on a real corporate acquisition, the environment included approximately 4,000 anonymized files. AI models had to locate relevant documents, analyze them, and correct errors in existing records. Despite these efforts, many models failed to complete complex tasks fully. In a profitability analysis task, none of the 25 runs achieved complete success, with the best run passing only 8 out of 11 verification checks. In legal tasks, four models attempted to modify investment contracts 20 times but failed to meet completion criteria, often missing contractual loopholes or adding conflicting terms.
The developers also highlighted issues with scoring in other benchmarks, claiming that about 50 percent of tasks in a quality verification system had errors. They stated that AI could be penalized for correct solutions if they did not follow a specific format. GDPx was designed to verify both the tasks and the scoring criteria to prevent such errors. The results will be used to predict revenue for OpenAI and Anthropic in 2027 and 2028, focusing on sectors where automation yields the highest economic return.
Limitations of the study include the use of different agent execution environments for each model and safety restrictions that limited some cybersecurity and biology tasks. Some task details and scoring criteria remain unpublished, making independent replication difficult. The release of GDPx signals a shift in AI evaluation toward measuring specific industrial expertise and economic utility rather than broad cognitive ability.