Methodology

The Intelligence Index

The AI Leaderboard uses the verified AI Leaderboard Intelligence Index v1.0 as the primary ranking mechanism. This composite index provides a robust, holistic view of an AI model's capability across diverse domains.

The index incorporates 10 different evaluations. Each evaluation is weighted according to its predictive power, difficulty, and resistance to overfitting or "gaming."

Index Composition (v4.3)

  • Agents30%
  • - AIL-Briefcase15%
  • - GDPval-AIL v210%
  • - AutomationBench-AIL5%
  • Coding20%
  • - Terminal-Bench v4.010%
  • - SciCode10%
  • General30%
  • - AIL-Omniscience Accuracy10%
  • - AIL-Omniscience Non-hallucination5%
  • - GDP.pdf10%
  • - AIL-LCR v1.15%
  • Scientific Reasoning20%
  • - Humanity's Last Exam10%
  • - CritPt10%

Cost and Speed

Cost per Task is calculated based on a standardized 1,000 input tokens and 1,000 output tokens.Speed is measured in median output tokens per second, testing independent API instances where available.

Data Integrity

This platform strictly preserves source accuracy. If a specific model variant (e.g., max vs high effort) scores differently, they are tracked independently. Results are clearly attributed to their source and evaluation date.