Artificial Analysis Intelligence Index v4.3.2

What it measures

Artificial Analysis is an independent evaluator. It runs 10 evaluations itself under one harness and blends them, by four weighted categories, into a single 0–100 index; higher means stronger overall.

The current v4.3.2 leans toward agents doing real work: Agents carry 30% (AA-Briefcase v1.1, GDPval-AA v2.1 and AutomationBench-AA, which require file deliverables, workplace tasks or SaaS workflows driven through APIs); Coding 20% (Terminal-Bench 4.0 and SciCode); Scientific Reasoning 20% (Humanity’s Last Exam and the physics-research benchmark CritPt); and General 30% (AA-Omniscience for knowledge and hallucination, GDP.pdf for long-PDF reasoning, and AA-LCR v1.1 for long-context reasoning). It answers “roughly where does this model sit?”

quickly, but it cannot tell you about any single skill. Each reasoning-effort setting (low / high / max) of a model is scored as its own entry.

Reading the scores

a relative 0–100 index, only meaningful among models scored in the same period on the same version. V4.3 replaced 𝜏³-Banking with AutomationBench-AA and Terminal-Bench 2.1 with the harder 4.0, and v4.3.2 then re-anchored the GDPval-AA v2.1 Elo, so scores should not be compared directly with older versions. This page includes every entry scoring 45 or above on the default (Current) leaderboard as of October 16, 2026; Anthropic entries that AA labels “with fallback” include a backup model after refusals, and are named here by reasoning effort.

SourceArtificial Analysis

Read the official Intelligence Index methodology and version history