SemiAnalysis Introduces GDPx: A New Benchmark for AI Agent Performance in Professional Work

SemiAnalysis has released GDPx, a new benchmark for AI agents in simulated professional work environments, revealing that no model scored above 12 percent across 206 tasks in eleven domains such as chip design, biology, and law. The benchmark, published in early November 2026, highlights significant limitations in current AI capabilities for complex professional work. For broader context, explore our AI Tools Pricing.

Understanding the GDPx Benchmark

The GDPx benchmark comprises 206 distinct tasks spanning eleven professional domains. These domains include specialized fields such as chip design, biology, quantitative trading, cybersecurity, and law. The design of GDPx focuses on replicating scenarios that skilled professionals might encounter, some of which could take up to eight hours to complete.

SemiAnalysis developed GDPx to address perceived shortcomings in other benchmarks, such as OpenAI's GDPval, which SemiAnalysis suggests lacks realistic work environment simulations. The GDPx benchmark is intended to be refreshed every three to six months to keep pace with rapid advancements in AI technology.

Initial Performance of Leading AI Models

The inaugural run of GDPx revealed substantial limitations in current AI agent capabilities. Anthropic's Claude Opus 5.5 emerged as the top performer, achieving a score of 12 percent. Google's Gemini 4 Carbon followed with 10.9 percent, and OpenAI's GPT-6 Astra scored 9.6 percent.

A notable finding was the complete failure of any model to fully complete tasks within the finance or legal domains. This suggests that tasks requiring intricate professional software interaction or extensive data room searches remain largely beyond the current scope of AI agents.

Domain-Specific Performance Insights

While overall scores were low, some models showed stronger performance in specific areas when certain safety filters were adjusted. For instance, when safety-filtered domains were removed, GPT-6 Astra's score increased to 11.8 percent. Meta's Muse Wonder Full demonstrated a more significant capability in cybersecurity, achieving 35.2 percent in that specific domain within the GDPx benchmark.

Criticisms and Future Outlook

Despite its ambitious scope, the GDPx benchmark has faced some criticism. Concerns have been raised regarding its reliance on mostly private tasks, the use of partial large language model (LLM) grading, and the dependency on provider-specific agent software. These factors could potentially influence the generalizability and transparency of the benchmark's results.

The low scores across the board highlight the significant gap between current AI agent capabilities and the demands of complex professional work. The benchmark's findings suggest that while AI models are advancing, they are still far from autonomously handling many real-world, multi-step professional tasks that require nuanced understanding and interaction with specialized tools and large datasets.

Conclusion

The introduction of SemiAnalysis's GDPx benchmark provides a new, more rigorous lens through which to view the practical capabilities of AI agents. The initial results underscore that even leading models have considerable room for improvement in handling complex professional tasks, particularly those involving specialized software and extensive data analysis. As SemiAnalysis plans to update the benchmark regularly, it will be a key indicator for tracking the progress of AI agents in real-world applications.

Sources

• SemiAnalysis-InferenceX/AGENTS.md at main · setyolegowo/SemiAnalysis-InferenceX · GitHub
• semianalysisai (SemiAnalysis) – Datasets
• GDPx: Measuring Frontier Model Capabilities Across the Most Valuable Domains