Artificial Analysis launches Intelligence Index v5: 60% private test sets, Claude Opus 5.5 leads at 58 and the open-weights gap widens to 12 points

Background and the problem

Public benchmarks are increasingly saturated or contaminated by training data, so a single score says less and less about real capability. Artificial Analysis is an independent model evaluator, and its Intelligence Index is a widely used composite measure. Version 5, released Oct 28, is its biggest revision since v4 in January, and the key change is that 60% of the index now comes from evaluations with private test sets or held-out answers, which limits contamination and targeted tuning.

How the index is built

v5 has nine evaluations in four categories weighted 25% each: Knowledge Work (AA-Briefcase 15%, AutomationBench-AA 5%, GDP.pdf 5%), Coding (Cognition's newly launched private FrontierCode 2.0 15%, Terminal-Bench 4.0 10%), Science (Terminal-Bench-Science 15%, a private 40-question hard subset of CritPt v2 10%) and General (AA-Omniscience v1.1 15%, HLE-Diamond 10%). HLE-Diamond is the refined 1,000-question Humanity's Last Exam from CAIS and Scale AI; GDPval-AA v2.1, SciCode and AA-LCR were removed. Agentic evaluations run in Artificial Analysis's own Stirrup harness. It also launched Progress Index v1, which places models released since November 2022 on one scale and builds on learnings from Epoch AI's capabilities index.

Results (as reported by Artificial Analysis)

Claude Opus 5.5 (max) leads at 58, followed by GPT-6 Astra (max) at 57, GPT-6.1 Sol (max) at 55 and Gemini 4 Argon (high) at 50. No model leads all nine evaluations: Opus 5.5 leads FrontierCode 2.0 (46%), HLE-Diamond (68%) and AA-Omniscience; GPT-6 Astra leads Terminal-Bench-Science (63%) and CritPt (72%); Kimi K3.1 Pro leads Terminal-Bench 4.0 (67%); Gemini 4 Argon leads AutomationBench-AA (78%). Among open weights, Kimi K3.1 Pro scores 46 and released-weight leaders DeepSeek V4.1 Flash 1017 and GLM-5.3 score 32, widening the gap from 3 points in v4.3 to 12. On cost, Opus 5.5 (max) is $10.04 per task and GPT-6.1 Sol (max) $1.27.

What developers can do

Don't pick on the headline number alone: for coding agents, look at FrontierCode 2.0 and Terminal-Bench 4.0; for research or data workflows, look at Terminal-Bench-Science. v5 shows that effort settings move both score and cost a lot (Opus 5.5 scores 35 at low effort and 58 at max for 10x the cost), so route by task difficulty. v5 scores are not directly comparable with v4.3, so older comparisons need a fresh look. Full rankings and methodology are on the Artificial Analysis site.