A new artificial intelligence (AI) benchmark has revealed that even the world’s most advanced AI models struggle to complete difficult professional tasks. The top model scored just 12% on this particular benchmark, highlighting how difficult it remains for AI agents to complete specialized, multi-step work without human assistance.
The findings were released on November 2, 2026, with the introduction of GDPx, a new benchmark designed to measure how well advanced AI models perform in real-world professional environments. GDPx focuses on demanding, interactive assignments that require AI agents to gather information, use software tools, solve problems, make decisions, and complete several connected steps.
The benchmark was developed by AI and semiconductor research firm SemiAnalysis in collaboration with 14 specialized data partners. It includes 206 tasks across 11 specialist domains: software engineering, cybersecurity, finance, legal work, biology, chip design, hardware design, materials science, quantitative trading, recursive self-improvement, and AI alignment.
The official GDPx leaderboard shows how leading AI models performed on these demanding professional tasks. Anthropic’s Claude Opus 5.5 finished first with an overall score of 12%. Google’s Gemini 4 Carbon followed with 10.9%, while OpenAI’s GPT-6 Astra ranked third with 9.6%. Other advanced models also recorded relatively low scores, showing how challenging these specialized tasks remain for current AI systems.
These results do not mean AI models fail at 88% of every task people give them. GDPx focuses on unusually difficult professional assignments, and its scores combine different types of task evaluations. However, the low results suggest that even powerful AI systems still face major limitations when asked to perform complex work independently.
To calculate the overall scores, the researchers gave equal weight to each of the 11 domains and evaluated model performance across three attempts per task. The benchmark uses both pass-or-fail grading and partial-credit scoring, depending on the assignment. Most tasks had a default time limit of eight hours, and models used their own agent tools when available.
The benchmark places particular importance on AI agents, which can use software tools and complete multi-step tasks instead of simply responding to questions. An AI agent may inspect documents, use software tools, analyze information, write code, make decisions, and complete several stages of a project. Such systems are becoming increasingly important as technology companies develop AI tools designed to support businesses and professionals.
GDPx attempts to test these capabilities through challenging assignments that resemble actual workplace responsibilities. For example, a software engineering task may require an AI model to understand an unfamiliar program, identify a problem, make changes, and confirm that the software continues working correctly. In one GDPx finance setting, AI agents must examine thousands of anonymized documents from a private-equity transaction before completing complex financial or accounting assignments.
Compared with OpenAI’s earlier GDPval benchmark, which primarily tests one-shot professional deliverables, GDPx places more emphasis on interactive tasks where AI agents must investigate information and work through several steps.
Finance and legal tasks proved especially difficult, with almost all models scoring zero under strict pass-or-fail grading. Some models completed parts of these assignments correctly, but failed to satisfy all the requirements needed for full credit. These assignments required detailed understanding, careful analysis, and accurate results across several connected requirements. Even when AI systems completed parts of the work correctly, they often struggled to meet every condition needed for a successful outcome.
The findings also showed that different AI models performed better in different fields. Claude Opus 5.5 led in areas such as chip design, hardware design, and quantitative trading. Gemini 4 Carbon performed strongly in biology and software engineering, while GPT-6 Astra showed stronger results in certain AI improvement and materials science tasks.
Another important finding involved AI safety systems. SemiAnalysis said safety filters affected the results of some models, including Claude Opus 5.5 and GPT-6 Astra, particularly in cybersecurity, biology, and AI alignment tasks. This suggests that benchmark results can reflect both technical capabilities and the limits placed on models to prevent potentially harmful behavior.
For businesses, the results highlight the importance of human supervision when using AI for important operations. AI tools can already help developers write code, support researchers, review information, and improve productivity. However, allowing these systems to handle complex financial, legal, engineering, or security work without human oversight, careful testing, and appropriate safeguards could create serious risks.
The findings also matter for businesses investing in AI tools and automation. Choosing a model based only on its overall ranking may overlook important differences in performance across individual fields. Companies should consider task accuracy, operating costs, safety requirements, and the need for human supervision before giving AI agents more responsibility.
At the same time, the results should not be viewed as proof that AI development is slowing down. The benchmark was intentionally designed to test difficult capabilities that existing evaluations may not measure effectively. As AI models become more advanced, their performance on these tasks could improve.
The researchers plan to update GDPx regularly, with new versions expected every three to six months. These updates could help track whether improvements in AI models lead to better results on complicated professional assignments.
The latest findings show that artificial intelligence has made significant progress but still has considerable room for improvement. For businesses and everyday users, the message is clear: AI can be a powerful assistant, but difficult and high-stakes tasks still require careful human oversight. As technology companies develop more capable AI agents, concerns about AI loss-of-control incidents underline why reliability, accuracy, and responsible decision-making remain essential measures of progress.
