SemiAnalysis releases GDPx benchmark for frontier AI models

The results show that even the best-performing models score only 12% on the full benchmark, indicating significant room for improvement. OpenAI and Anthropic models generally lead the rankings, with Opus 5.5 and GPT-6 Astra performing strongly, though their scores are sometimes lowered by safety classifiers. Gemini 4 Carbon is noted as a strong competitor, particularly in biology and software engineering. The authors highlight that safety classifiers significantly impact scores in domains like cyber and alignment, and that current models struggle heavily with complex finance and legal tasks, where many scored zero.

SemiAnalysis argues that existing benchmarks are inadequate for evaluating today's agents because they lack realistic work environments and often contain unfair grading criteria. By partnering with domain experts, GDPx aims to provide a more accurate picture of where frontier models can actually be deployed at high ROI. The benchmark is available for public exploration, and the authors suggest that tracking progress in these specific high-impact domains is more useful than broad GDP-based metrics.