Artificial Analysis Intelligence Index explained: what is in the score
The Artificial Analysis Intelligence Index averages nine evaluations in four equal categories. What v5.0.1 contains, how it is run and what it leaves out.
The Artificial Analysis Intelligence Index is a single score for a language model, built by averaging results on a fixed set of evaluations that Artificial Analysis runs itself. Artificial Analysis calls it a synthesis of overall model capability across knowledge work, coding, science and general knowledge. It is the number you see attached to many model launches, so it helps to know what it contains.
The index is versioned and its contents change. As of November 2026, the methodology page describes v5.0.1, which incorporates nine evaluations. Older write-ups that list a different mix are describing earlier versions.
What v5.0.1 contains
The methodology page says the index is a weighted average over four categories, each worth 25%.
| Category (25% each) | Evaluation | Weight in the index |
|---|---|---|
| Knowledge Work | AA-Briefcase v1.1 | 15% |
| Knowledge Work | AutomationBench-AA | 5% |
| Knowledge Work | GDP.pdf | 5% |
| Coding | FrontierCode-AA 2.0 | 15% |
| Coding | Terminal-Bench 4.0 | 10% |
| Science | Terminal-Bench-Science 0.1.0 | 15% |
| Science | CritPt v2 (Hard Subset) | 10% |
| General | AA-Omniscience v1.1 | 15% (10% accuracy, 5% for 1 minus hallucination rate) |
| General | HLE-Diamond | 10% |
HLE-Diamond is, per the methodology page, a refined 1,000-question version of Humanity’s Last Exam from the Center for AI Safety and Scale AI, split into 500 reasoning and 500 knowledge questions. The Terminal-Bench entries relate to Terminal-Bench. The AA-Omniscience weight is split so that a model is rewarded for answering correctly and also for not hallucinating.
How the runs are set up
- Scoring. Artificial Analysis generally uses pass@1, aggregated across repeats, where the model must be right on its first attempt.
- Settings. Temperature is 0 for non-reasoning models and 0.6 for reasoning models, unless the model’s lab recommends another value. Non-reasoning models get up to 16,384 output tokens. See LLM temperature for what the setting does.
- Failures. API calls are retried up to 30 times, and questions that fail every retry are reviewed manually.
- Elo-based evaluation. For AA-Briefcase, the page says the Elo rating is frozen when a model is added and normalized as clamp((Elo - 500) / 2000), with the scale anchored to GPT-5.5 (medium) at 1000. Artificial Analysis may update the reference parameters later.
How the index has changed
The methodology page records several changes, which is the main reason to be careful with old scores.
- GDPval-AA was a constituent up to and including v4.3.2 and was removed in v5.0. It is still run as a standalone evaluation. See GDPval for the underlying benchmark.
- IFBench left the index in v4.1.
- Terminal-Bench 2.1 was superseded by Terminal-Bench 4.0 in v4.3.
- AA-Omniscience v1.1 has 3,000 questions, down from 6,000, after saturated and flawed questions were removed. Scores are not directly comparable with v1.
Because of this, a number quoted without its version is hard to interpret. Third-party pages that describe different category weights are describing an older version.
What the index leaves out
| Limit | What the methodology says |
|---|---|
| Modality and language | The suite is primarily text-based and English-language. Images appear in a few evaluations, such as HLE-Diamond image questions, but visual, speech and multilingual performance are benchmarked separately. |
| Other capabilities | Extra evaluations for multilingual, visual and mathematical work are reported separately and do not enter the score. |
| Fit to your use case | Artificial Analysis says that, like all evaluation metrics, the index has limitations and may not apply directly to every use case. |
| Comparisons | Artificial Analysis states it is confident the index is a more useful comparison than any other metric. That is the publisher’s claim, not an independent finding. |
Using it to choose a model
Treat the index as a way to build a shortlist. A composite hides which category drove a score, so read the per-evaluation results when your work is mostly coding or mostly document-heavy knowledge work. For preference-based rankings, see LMArena. For benchmarks that grade multi-step agent tasks, see agent benchmarks, and for the broader method, LLM evaluation.
Testing candidates on your own data
Public indexes do not know your prompts, tools or output format. HoneyHive lets you run candidate models over a fixed set of your own inputs with datasets and experiments, and score the results with your own evaluators.
FAQ
What is the Artificial Analysis Intelligence Index?
A composite score for language models that Artificial Analysis computes itself. Its methodology page describes it as a weighted average of evaluations across knowledge work, coding, science and general knowledge.
What is in the Intelligence Index?
As of v5.0.1 (November 2026), nine evaluations: AA-Briefcase, AutomationBench-AA, GDP.pdf, FrontierCode-AA, Terminal-Bench, Terminal-Bench-Science, CritPt, AA-Omniscience and HLE-Diamond, per the methodology page.
Do Intelligence Index scores from different versions compare?
Not safely. The composition has changed between versions, and Artificial Analysis notes that scores on a revised evaluation such as AA-Omniscience v1.1 are not directly comparable with the earlier version.
Does the Intelligence Index cover images, speech or other languages?
No. It is primarily text-based and English-language. Visual reasoning, speech and multilingual performance are benchmarked separately.
Can the Intelligence Index tell me which model to use?
It is a shortlisting signal. Artificial Analysis itself says the index may not apply directly to every use case, so test candidates on your own data.