What it is
The Artificial Analysis Intelligence Index is a single composite score meant to summarize how capable a given AI model is. It is produced by Artificial Analysis, a company that independently benchmarks AI models, cloud infrastructure, and chips, and that markets itself on having no commercial ties to the model developers it rates. The index combines nine separate evaluations into one number, weighted equally across four domains: knowledge work (25%), coding (25%), science (25%), and general reasoning (25%). Every model is tested under identical conditions — same prompting strategy, same temperature settings, same scoring method — so that, in principle, scores are comparable across providers from OpenAI, Google, Anthropic, Mistral, Moonshot AI, and others.
The index has gone through several major revisions as developers have gotten better at tuning their models directly to known public benchmarks, a problem widely called “benchmark contamination” or “saturation.” The newest revision, version 5.0, was published on 27 October 2026. Its headline change is a shift toward secrecy: for the first time, a majority of the index — 60% — is built on evaluations with private test sets or held-out answers that model developers cannot see or train against. The new version adds two benchmarks (FrontierCode 2.0 for real-world coding tasks, and Terminal-Bench-Science for agentic scientific research) and upgrades three others, including replacing the well-known public “Humanity’s Last Exam” with a harder, partly-private successor called HLE-Diamond.
Under v5.0, Google’s Gemini 4 Carbon took the top spot with a score of 59, edging out Claude Opus 5.5, which had led the prior version with 58. Mistral’s new Large 4 model fell sharply, from 38 points under the old methodology to 25 under the new one — not because the model changed, but because the test composition did.
Why it matters for AI governance and narratives
Benchmark leaderboards like this one have become a de facto, privately-owned infrastructure for how the AI industry, journalists, and policymakers talk about progress and hierarchy. A single number lets a company claim “state of the art,” lets a journalist write a clean lede, and lets a regulator or investor make a rough capability comparison without reading any technical documentation. That convenience obscures how contestable the number is: a one-point methodology change can reorder the field, as the Mistral Large 4 case shows, and once the majority of underlying test questions are secret, outside researchers lose the ability to independently verify the ranking at all. Coverage following the v5.0 release noted this tension directly — that a benchmark designed to resist gaming by developers also, as a side effect, resists scrutiny by everyone else, including the press that cites it as ground truth.
This matters for the observatory’s purposes because “Index v5.0 crowns a new leader” is not a neutral finding — it is the output of one company’s test design, published with enough authority that it shapes which model gets called the frontier model in the same week trial dates are being set in the Apple–OpenAI trade-secret litigation. Measurement itself is a site of framing contest, not a referee standing outside it.
Key facts and dates
Artificial Analysis was started in 2023 by co-founders Micah Hill-Smith (CEO) and George Cameron (CPO), and launched publicly in 2024. The company raised early funding from AI Grant, the seed program run by former GitHub CEO Nat Friedman and investor Daniel Gross, who provided both capital and cloud compute credits. The company’s stated core commitment is independence from the model developers it benchmarks — it takes no funding or sponsorship from the labs whose models appear on its leaderboard, according to its own description of its methodology.
The Intelligence Index has iterated through multiple versions as public test sets became saturated; v4.2 was described by the company as an acceleration of the v5 roadmap, specifically to push toward long-horizon, real-world tasks and a greater share of private test data. Version 5.0, released 27 October 2026, is the first version where private evaluations form a majority of the score. One side effect flagged in coverage: models strong on the index’s few remaining public benchmarks — for instance, Moonshot AI’s Kimi K3.1 Pro, which led every proprietary model on the public agentic-coding test Terminal-Bench 4.0 — can nonetheless rank far down the composite (46th, down from 4th in the prior version) once private, harder-to-optimize-for tests dominate the weighting.