How to Read the Artificial Analysis Intelligence Index v5
Nine weighted evaluations, with cost per task for the top models.
Artificial Analysis Intelligence Index v5 is a weighted average of nine evaluations in four categories at 25% each. Sixty percent of it rests on private test sets or held-out answers, and Artificial Analysis says its scores are not meant for comparison across versions. Read a v5 score as a shortlist signal, read cost per task as the budget signal, and use the Progress Index only to compare models released years apart.
At Layer3Labs, we build custom agents and enterprise integrations for small and mid-sized businesses.
Published on October 27, 2026, by Artificial Analysis, v5 marks the largest update to the index since v4 launched in January.
How to Read an Intelligence Index v5 Score
An Artificial Analysis Intelligence Index score represents a blended composite across four categories: Knowledge Work, Coding, Science and General. Each category carries exactly 25% of the total index weight, split across nine specific evaluations.
In our internal check of the published figures, we summed the weights and compared score gaps with cost per task. The four categories each sum to 25%. Across nine evaluations, weights total 100% (15 + 5 + 5 + 15 + 10 + 15 + 10 + 15 + 10), matching published figures.
| Category | Weight | Evaluations Included | Evaluation Weights |
|---|---|---|---|
| Knowledge Work | 25% | AA-Briefcase, AutomationBench-AA, GDP.pdf | AA-Briefcase: 15%, AutomationBench-AA: 5%, GDP.pdf: 5% |
| Coding | 25% | FrontierCode 2.0, Terminal-Bench 4.0 | FrontierCode 2.0: 15%, Terminal-Bench 4.0: 10% |
| Science | 25% | Terminal-Bench-Science, CritPt v2 hard subset | Terminal-Bench-Science: 15%, CritPt v2 hard subset: 10% |
| General | 25% | AA-Omniscience v1.1, HLE-Diamond | AA-Omniscience v1.1: 15% (10% accuracy, 5% non-hallucination), HLE-Diamond: 10% |
Check whether your daily workload matches this even split before you trust the average. If your internal applications focus entirely on document processing or customer operations, index weights will not match your workload. Coding carries 25% of the total score, and Science carries another 25%.
What Changed From v4
The primary structural shift in v5 is that 60% of the index now relies on private test sets or held-out answers. This marks the first edition where the majority of the benchmark uses non-public datasets. Artificial Analysis introduced two new benchmarks, updated three existing suites, retained four, and eliminated three older evaluations.
FrontierCode 2.0, developed by Cognition, provides a private coding benchmark containing 132 tasks across open-source repositories. Each task runs three times in Stirrup, the Artificial Analysis agentic harness, and is graded on whether a repository maintainer would merge the change. Any failed blocking check scores zero. The second addition, Terminal-Bench-Science, evaluates 70 agentic scientific research tasks executed in a terminal environment.
Three evaluations received substantive overhauls. HLE-Diamond replaces Humanity's Last Exam, presenting 1,000 questions split evenly between reasoning and knowledge. CritPt v2 hard subset introduces a private 40-question subset revised by the CritPt authors and executed using a code-execution tool. AA-Omniscience v1.1 removed saturated and public-set questions, added fresh items, and corrected ambiguous questions. On AA-Omniscience v1.1, scores mostly align with v1, and models average 2 points lower (median 0.6).
Artificial Analysis also removed GDPval-AA v2.1, SciCode, and AA-LCR v1.1. Four tests were retained from previous iterations: AA-Briefcase, AutomationBench-AA, GDP.pdf, and Terminal-Bench 4.0. Artificial Analysis says Intelligence Index scores are not meant for long-run comparison because the evaluations change across versions, so do not set a v5 score next to a v4 score.
Reading Score Against Cost per Task
Artificial Analysis reports cost per Intelligence Index task for the ten highest-scoring models, and that cost varies by more than 13x. The lowest-cost option in the top ten is GPT-6.1 Terra (max) at $0.96 per task, while Claude Fable 5.1 (max) reaches $13.04 per task.
| Model | Intelligence Index v5 Score | Cost per Task |
|---|---|---|
| Claude Opus 5.5 (max) | 58 | $10.04 |
| GPT-6 Astra (max) | 57 | $6.24 |
| GPT-6.1 Sol (max) | 55 | $1.27 |
In our internal check of the published figures, we compared the score gaps against cost ratios for the top performers. Claude Opus 5.5 (max) scores 58, which is 3 points above GPT-6.1 Sol (max) at 55. However, Claude Opus 5.5 costs $10.04 per task compared to $1.27 for GPT-6.1 Sol. That is a 7.9x cost ratio, meaning the Intelligence Index run costs almost eight times as much per task for a 3-point composite gain.
Similarly, GPT-6 Astra (max) scores 57, putting it 2 points above GPT-6.1 Sol (max), but costs $6.24 per task. That represents a 4.9x cost ratio ($6.24 / $1.27) for a 2-point difference on the composite index. You can review broader operational pricing patterns in our guide to AI model pricing.
Effort settings change both score and cost: Claude Opus 5.5 scores 35 at low effort ($0.98 per task) and 58 at max ($10.04). Claude Sonnet 5.5 spans 29 points from low to max effort, up from 20 in v4.3. Meanwhile, the spread for GPT-6.1 Sol doubled from 10 points to 20 points.
Median output tokens per task rose by two-thirds from v4.3 to v5, from 35k to 59k. Four agentic evaluations (Terminal-Bench-Science, Terminal-Bench 4.0, AA-Briefcase, and FrontierCode 2.0) account for 55% of the total index weight (15 + 10 + 15 + 15 = 55%). However, they consume a median of about 90% of all generated output tokens.
The Progress Index v1
To evaluate historical performance trends without the distortion of shifting benchmark suites, Artificial Analysis introduced the Progress Index v1.0 in October 2026. See details on the Artificial Analysis methodology page.
While the Artificial Analysis Intelligence Index remains the primary measure of current capabilities, the Progress Index is designed for long-run trend tracking across models released years apart.
It calculates scores using item response theory. Models do not need to have run the exact same evaluations, because shared benchmarks link different model generations onto a common scale. To be included, a benchmark must have results from at least 4 models, and a model must have results across at least 4 benchmarks.
Anchors are o3-mini (high) at 220 and GPT-5.5 (xhigh) at 320. The methodology relies on Artificial Analysis's own evaluations. The published documentation does not assign a specific unit of measurement to the scale, nor does it list every covered model.
When the Index Misleads You
An aggregate score hides category trade-offs that can mislead technical decision-makers. Benchmark leaders differ from the overall top spot. On FrontierCode 2.0, Claude Opus 5.5 (max) leads with 46%. On Terminal-Bench-Science, GPT-6 Astra (max) leads with 63%. On AutomationBench-AA, Gemini 4 Argon leads with 78%, and on the CritPt v2 hard subset, GPT-6 Astra leads at 72% under high effort.
Even within a single test, skill profiles vary. On HLE-Diamond, Claude Opus 5.5 and GPT-6 Astra both score 78% on reasoning questions, but score 58% and 50% respectively on knowledge questions. If your work leans on knowledge questions, the composite HLE-Diamond score hides that gap. To evaluate how autonomous agent workflows behave under similar conditions, see our analysis of AI agent benchmarks.
Open-weights performance also shows a widening gap at the frontier. Kimi K3.1 Pro reaches a score of 46 as the highest-ranking open-weights model, though its weights have not yet been released. Among models with currently released weights, DeepSeek V4.1 Flash 1017 (max) and GLM-5.3 (max) lead at 32 points. Artificial Analysis reports that the gap between open-weights models and the leader expanded to 12 points in v5, up from 3 points in v4.3. For context on broader performance baselines, consult our overview of AI model benchmarks.
The published sources do not show how well index rank predicts results on any particular business task. For practical deployment, use the Artificial Analysis Intelligence Index to build a shortlist of two or three models. Use cost per task to define your operational budget, and run 20 of your own real tasks before committing. If one category is most of your work, read that category's leader. This recommendation would change if Artificial Analysis alters its category weights or benchmark mix again.
Frequently Asked Questions
- The Artificial Analysis Intelligence Index is a composite benchmark that is the primary measure of current model intelligence, per Artificial Analysis. Nine evaluations span four evenly weighted categories.
- Artificial Analysis says the Intelligence Index's evaluations change across versions, so its scores are not meant for long-run comparison. Version 5 added two evaluations, upgraded three, retained four and removed three, and 60% of the index now uses private test sets or held-out answers.
- It tracks model progress since November 2022. It uses item response theory to place models on a shared scale anchored by o3-mini (high) at 220 and GPT-5.5 (xhigh) at 320. That lets it compare models evaluated on different benchmark suites over time.
- Claude Opus 5.5 (max) holds the highest composite score on Intelligence Index v5 with 58 points. GPT-6 Astra (max) follows at 57 points, and GPT-6.1 Sol (max) scores 55 points.
- No. The cost per task among top models varies by more than 13x. For example, Claude Opus 5.5 scores 3 points higher than GPT-6.1 Sol (58 versus 55) but costs $10.04 per Intelligence Index task compared to $1.27. That represents a 7.9x cost ratio for a 3-point composite score gain.
Evaluating AI Models for Your Team?
Layer3Labs helps businesses scope custom agents and integrate automated workflows around real operating constraints.
Book a Consultation