Artificial Analysis Intelligence Index
Most fields on a gallery card describe the model itself, such as its attention type, layer mix, or KV-cache size. I added the Artificial Analysis Intelligence Index to provide a separate performance reference. It answers a different question. How did the released model score on a shared set of evaluations?
The card shows this information in two lines. Total score is copied from the matching Artificial Analysis model page. Profile is my compact diagnostic view across Agents, Coding, General, and Scientific Reasoning.
There is an important distinction between them. The total is the official Intelligence Index. The profile is specific to this gallery and should not be used to reconstruct the total.
Current method
Intelligence Index v4.3.2, primarily text-based and English-language
Category weights
Agents 30%, Coding 20%, Scientific Reasoning 20%, General 30%
Gallery data snapshot
2026-10-17
What the two fields mean
Total score is the overall Intelligence Index shown on the matching Artificial Analysis model page. If that page reports an estimate while an independent evaluation is pending, the gallery records the estimate.
Profile is a second view that I added for quick comparisons. Each category uses the v4.3.2 components below, with weights normalized within that category to a 0–100 scale. A category shows N/A if any required component is unavailable; a published zero remains zero.
For example, the gallery snapshot for Nemotron 3.5 Super VL records a total of 26.5 alongside General 32.6, Scientific 20.6, Coding 23.9, and Agents 26.0. The total is copied directly from Artificial Analysis. It is never recalculated from the profile.
Agents uses the published evaluation Elo ratings, normalized with clamp((Elo − 500) / 2000, 0, 1). Artificial Analysis freezes Elo inputs for its official total, so a profile derived from published ratings can differ slightly, even before rounding.
I keep the profile because two models with similar totals can still look quite different on coding or agent tasks. It is a directional comparison rather than another official Artificial Analysis index.
What goes into v4.3.2
The current index groups ten evaluations into four categories:
- Agents, 30%. AA-Briefcase v1.1 contributes 15%, GDPval-AA v2.1 contributes 10%, and AutomationBench-AA contributes 5%.
- Coding, 20%. Terminal-Bench 4.0 and SciCode each contribute 10%.
- General, 30%. AA-Omniscience supplies 10% accuracy and 5% non-hallucination, GDP.pdf contributes 10%, and AA-LCR v1.1 contributes 5%.
- Scientific Reasoning, 20%. Humanity’s Last Exam and CritPt each contribute 10%.
Earlier gallery snapshots used a different profile recipe, including IFBench and GPQA Diamond. The October 17 refresh updates the profile along with the total. Compare scores within the same snapshot, rather than treating changes across methodology versions as changes in model capability.
Why the snapshot date matters
Artificial Analysis revises both benchmark composition and model results. For that reason, the gallery stores the date of its data pull. The current snapshot on this page is 2026-10-17.
The date matters in model comparisons. Scores collected under different index versions may reflect changes in the evaluation suite as well as changes in the models. When a matching model page or a required field is unavailable, the gallery displays N/A instead of filling the gap with an estimate.
What the score can tell us
I use the index as a broad check on a model release. A score difference cannot identify which design choice caused it. Training data, post-training, tool use, reasoning settings, and test-time compute can all influence the result alongside architecture.
The suite primarily covers text-based, English-language tasks. It does not replace separate assessments of vision, speech, or multilingual performance.
For architecture comparisons, I therefore use the index as supporting context. The architecture fields and the original technical report remain the better sources for understanding how a model is built.