SemiAnalysis has introduced the GDPx benchmark, designed to evaluate AI agents on tasks simulating knowledge-based work. According to The Decoder, no AI model tested managed to score above 12 percent, highlighting current limitations in AI performance on complex intellectual tasks.
Among the models assessed, Claude Opus 5.5 narrowly outperformed its closest competitor, Gemini 4 Carbon, though both remained well below the 12 percent threshold. This suggests that while some progress is evident, AI agents still face significant challenges in achieving higher capabilities in knowledge work simulations.
For Japanese markets, where automation and AI integration are rapidly advancing in sectors such as finance and technology, these findings underscore the importance of cautious optimism regarding AI’s near-term impact on complex decision-making and analysis.