A new benchmark called GDPx, developed by SemiAnalysis, evaluates AI models on complex, multi-step tasks lasting up to eight hours. The benchmark revealed that even the top-performing model, Claude Opus 5.5, only achieved a 12% success rate, with Gemini 4 Carbon and GPT-6 Astra scoring slightly lower. These results challenge the notion of fully autonomous AI agents, suggesting that current AI is more suited to augmenting human experts rather than replacing them, especially in fields like finance and law where models struggled significantly. AI
IMPACT Current AI models are not yet capable of autonomous expert work, indicating a need for AI roadmaps focused on augmenting human capabilities.
RANK_REASON New benchmark evaluating AI model performance on complex, long-duration tasks. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
