Artificial Analysis Revamps Its Smart Ranking List, With 40% of the Weight Now Based on Private Testing.

Insight Beating AI Flash News: Artificial Analysis has updated its Intelligence Index to version 4.2, designed to make its rankings more resistant to targeted score manipulation. The weight of private testing has been increased from 20% to 40%, preventing model developers from accessing the full test set in advance and reducing room for benchmark-specific optimization. The new version adds two test modules: AA-Briefcase and GDP.pdf. AA-Briefcase simulates real-world knowledge work projects spanning weeks, requiring models to handle numerous related tasks and thousands of documents. GDP.pdf uses 100 professional PDFs totaling 4,592 pages, challenging models to locate evidence across texts, tables, charts, and footnotes to answer questions. GPQA Diamond has been removed from the overall index. This graduate-level science question benchmark has reached near-saturation: GPT-6 Astra scored the highest at 96.3%, with Gemini 3.8 Flash at 95.3%, making it nearly impossible to distinguish between the top models. After recalculation, Claude Fable 5.1 retains the first place, followed by GPT-6 Astra. Astra notched a score of 55, 4 points higher than GPT-5.6 Sol. Artificial Analysis also noted that Astra uses fewer output tokens compared to other leading models in its tier.