Artificial Analysis Intelligence Index

How Artificial Analysis Intelligence Index works

The index aggregates separate evaluations into four categories. In v5, current since 27 October 2026, each carries 25%. Knowledge Work is AA-Briefcase at 15%, AutomationBench-AA at 5% and GDP.pdf at 5%. Coding is FrontierCode 2.0 at 15% and Terminal-Bench 4.0 at 10%. Science is Terminal-Bench-Science at 15% and the CritPt v2 hard subset at 10%. General is AA-Omniscience v1.1 at 15% (10% accuracy, 5% non-hallucination) and HLE-Diamond at 10%. About 60% of the v5 weighting now runs on private test sets or held-out answers, up from 45% in v4.3, which is the firm's defence against labs training on the test.

History and current status

Artificial Analysis has run the index since 2024, revising it whenever components saturate. The recent cadence is the thing worth knowing. v4.1 shifted the index toward agentic workloads. v4.2 landed on 4 September 2026, adding AA-Briefcase and GDP.pdf, retiring GPQA Diamond as saturated, and doubling the private-test weighting to 40%. v4.3 landed on 7 September 2026, three days later, upgrading Terminal-Bench from v2.1 to v4.0, replacing t3-Banking with AutomationBench-AA, and raising private weighting to 45%. v5 landed on 27 October 2026 as the largest rebuild since v4: it added FrontierCode 2.0 (Cognition's private, maintainer-graded coding set) and Terminal-Bench-Science, replaced Humanity's Last Exam with the 1,000-question HLE-Diamond, upgraded AA-Omniscience and CritPt, removed GDPval-AA, SciCode and AA-LCR, and moved every category to an equal 25%.

What the score does not tell you

It is a vendor-maintained composite, so its absolute values are a property of the current weighting rather than of the models. Because components are swapped and reweighted, scores do not compare across versions, and nothing in a quoted number tells you which version it came from. The firm also publishes several reasoning-effort variants per model, max, xhigh, high, medium and low, which produce materially different scores for the same model name, so a citation without the effort setting is ambiguous. Per-evaluation columns have been dropped from the public leaderboard at times, which makes it harder to check what is driving a score. Many leaderboard values carry a trailing asterisk for which no public legend is published.

This is the general case, not a quirk of one benchmark. For the four failure modes that decide whether any benchmark number means anything, read are AI benchmarks reliable, and for this row in particular, how benchmark contamination inflates scores.

Artificial Analysis Intelligence Index is not one of the 15 benchmarks scored in the benchmark trust ranking, which rates the most-quoted tests on those same four axes and sorts them most trustworthy first.

What Artificial Analysis Intelligence Index scores actually mean

The clearest way to read this index is to watch one model across versions. Under v4.1 the top of the board ran to about 66. Under v4.2, checked on 5 September 2026, Claude Fable 5.1 topped it at 57. Under v4.3, checked on 12 September 2026, the top read 53. Under v5, checked on 28 October 2026, Claude Opus 5.5 at max effort leads at 58, the same figure it held on v4.3.2, while Kimi K3.1 Pro fell from 55 to 46 and GLM-5.3 from 45 to 32. The models did not change; the index did. So a gap of a few points between two models on the same version is a real if modest difference, while the same gap between two versions is noise introduced by the scoring. Treat the index as an ordering within one published version and never as an absolute level.

Who reports Artificial Analysis Intelligence Index, and how to read it

Almost everybody, usually without the version. Labs cite it in launch posts when it flatters them, aggregators such as Price Per Token draw on it downstream, and press coverage reduces it to a single figure. That downstream reach is precisely why the version matters: a number republished from a v4.1-era article and set beside a v4.3 number implies a decline that never happened. When you see an index score quoted, the three things to ask for are the version, the reasoning-effort variant, and the date it was read.

When to weight Artificial Analysis Intelligence Index in a model choice

Use it as a first-pass filter when you need one yardstick across many models, which is what it is good at, and then check the component that matches your actual workload. If you are buying coding capability, FrontierCode 2.0 and Terminal-Bench 4.0 carry 25% between them and the underlying boards are more informative than the composite. If you are buying agentic knowledge work, AA-Briefcase, AutomationBench-AA and GDP.pdf carry 25%. Record the version, the effort variant and the read date alongside any number you store, because an index figure without those three is not reproducible and will silently rot inside whatever you put it in.