Artificial Analysis Intelligence Index

Where most benchmarks measure one thing, the Artificial Analysis Intelligence Index tries to compress many into one number. George Cameron and Micah Hill-Smith started the site in 2023 as a side project comparing model price and latency, after Hill-Smith found no independent source tracking which model was actually worth using for a given task. It grew into a continuously updated index that pairs a composite capability score with running price and throughput data pulled directly from providers’ APIs.

The index itself is a moving target by design. Version 4.1.1, current as of August 2026, weights nine separate evaluations into four categories — Agents (34%), Coding (24%), Scientific Reasoning (24%) and General (18%) — drawing on tests such as GDPval-AA, Terminal-Bench and Humanity’s Last Exam. Components and even their grading models are periodically swapped out as older ones saturate, which keeps the index responsive to real capability differences but means scores from different versions cannot be read as one continuous scale. The point was made plainly in September 2026, when version 4.3.2 rescored every model several points lower — GPT-6 Astra fell from 61 to 53 — and Claude Opus 5.5 (max) took the lead at 58, ahead of Claude Fable 5.1 and GPT-6 Astra at 53. A month later version 5.0 swapped in Terminal-Bench-Science, HLE-Diamond and FrontierCode 2.0 and dropped GDPval-AA, and when Gemini 4 Carbon was released on 28 October it took the lead at 59, a point ahead of Opus 5.5. Under v4.1.1, Claude Opus 5 (max) had led at 63; earlier in the summer, Zhipu’s open-weight GLM-5.2 had become the highest-scoring open model at 51, a result the company highlighted as evidence openly licensed models were closing on the closed frontier.

Because Artificial Analysis runs its own evaluations rather than relying on developers’ self-reported numbers, its index has become a common reference point in coverage of new releases, cited alongside — and sometimes instead of — a lab’s own benchmark claims. It is not peer-reviewed or academically governed, and its methodology choices, including which evaluations to include or retire, are set unilaterally by the company running it.