AInsights

AI Insights Analysis

Full Ranking Sources

Methodology

AInsights Index Methodology

The default AIndex ranking is calculated from five independent capability boards, with extra emphasis on Coding, Agentic/tool work, and Hard reasoning.

Capability Boards

AIndex now uses five capability boards instead of a regular-plus-bonus split. Scores are scaled to the current AA Intelligence maximum, and the boards are shown from highest to lowest weight.

Board Weight Selection logic
Coding 40 The largest board prioritizes real software engineering and difficult coding tasks: SWE-Bench Pro/Verified/Multilingual, AA Terminal-Bench v2.1, LiveCodeBench, Terminal-Bench Hard, and SciCode. Sparse benchmarks with fewer than four scored models remain source evidence but are not selected by the default matrix.
Agentic/tool work 24 Terminal-Bench v2.1, SWE-Bench, BrowseComp, HLE tools, MCP-Atlas, OSWorld, GDPval, Toolathlon, and AA-LCR capture tool, browser, terminal, desktop, and business workflow execution.
Hard reasoning 20 HLE, FrontierMath, CritPt, GPQA, AIME, and HMMT emphasize difficult problem solving. Sparse abstraction benchmarks with fewer than four scored models are not selected by the default matrix.
Knowledge/science 8 Omniscience, GPQA, HLE, lower-weight MMLU-Pro, MMLU, MMMU-Pro, and SciCode add scientific and factual breadth.
Instruction/context 8 IFBench, AA-LCR, CritPt, and CharXiv variants capture instruction following and context behavior.

A metric can appear in multiple boards because each board scores an independent capability dimension. For example, Terminal-Bench can reflect both code execution and tool workflow ability, but its weight inside each board only serves that board.

Metric Weights

Internal weights are normalized within each board. The matrix is ordered by board weight, and each board's metrics are ordered from highest to lowest internal weight.

Board AIndex weight Benchmark selection Metrics and internal weights
Coding 40 Real repository repair, terminal coding, contest-style generation, and harder coding/science tasks. SWE-Bench Pro 1.5LiveCodeBench 1.2SWE-bench Verified 1.1Terminal-Bench Hard 1.0Terminal-Bench v2.1 0.9SWE-bench Multilingual 0.8SciCode 0.8
Agentic/tool work 24 Tool use, browser research, terminal work, desktop tasks, business task completion, and long-context execution. SWE-Bench Pro 1.3BrowseComp 1.0HLE w/ tools 1.0MCP-Atlas Public 0.9OSWorld-Verified 0.8GDPval (wins or ties) 0.8Terminal-Bench Hard 0.8GDPval-AA Elo 0.7Toolathlon 0.6Terminal-Bench v2.1 0.5AA-LCR 0.5
Hard reasoning 20 High-difficulty general reasoning, math, science QA, contest math, and ARC-style abstraction. Humanity's Last Exam 1.3Humanity's Last Exam 1.2FrontierMath Tier 4 1.2CritPt 1.1FrontierMath Tier 1-3 1.0GPQA Diamond 0.9GPQA Diamond 0.8AIME 2025 0.7AIME 2025 0.7AIME 2026 0.3HMMT Feb 2026 0.3
Knowledge/science 8 Scientific and factual breadth across omniscience, GPQA, HLE, multilingual knowledge, multimodal reasoning, and code-adjacent science. AA-Omniscience Accuracy 1.0GPQA Diamond 0.9Humanity's Last Exam 0.8MMMLU 0.7MMMU-Pro 0.7SciCode 0.5MMLU-Pro 0.2
Instruction/context 8 Instruction following, long context, critique behavior, and chart/vision understanding with and without tools. IFBench 1.0AA-LCR 0.9CritPt 0.7CharXiv Reasoning w/ tools 0.6CharXiv Reasoning 0.5

Calculation Formula

For a metric m, the normalized metric value is ratio_m = max(raw_m, 0) / best_observed_m. Best observed means the highest score currently present in the dataset for that exact metric key.

For each board, AIndex first aggregates available metric ratios with the weighted geometric mean used by the code: exp(sum(weight_m * ln(1 + ratio_m)) / sum(available_weight_m)) - 1.

The board metric coverage adjustment is then applied: multiply the board value by (available_internal_weight / total_internal_weight) ^ exponent. The within-board coverage discount exponent is 0.10, including boards with only one available metric.

If a board has no available metrics, AIndex inserts a weak prior value of 0.34 for that board. If all five boards are missing, the model receives no default AIndex score.

The final board aggregation is another weighted geometric mean: AIndex = AA Intelligence max * (exp((40 * ln(1 + Coding) + 24 * ln(1 + Agentic) + 20 * ln(1 + HardReasoning) + 8 * ln(1 + KnowledgeScience) + 8 * ln(1 + InstructionContext)) / 100) - 1).

AIndex Calculation

  1. For each board metric, convert the raw score into a Best score ratio: model score divided by the highest score observed for that benchmark.
  2. Aggregate available metrics inside each board with a weighted geometric mean, then apply a within-board coverage discount using exponent 0.10, including when only one metric is present in that board.
  3. If an entire board is missing, use a weak prior of 0.34 for that board; if all boards are missing, no default AIndex is produced.
  4. Aggregate the five board values with board weights 40 / 24 / 20 / 8 / 8 using a geometric weighted mean.
  5. Scale the resulting ratio to the current AA Intelligence maximum.

Detailed Explanation

AIndex is designed to compare models across different benchmark scales without letting any single raw scoring convention dominate the final leaderboard. A percentage-style benchmark, an Elo-normalized field, and a long-context score can each have different ranges, so AIndex first places every active field on its own best-observed baseline.

The board structure is now the primary anchor. It deliberately gives the most weight to coding and agentic tool work, while keeping hard reasoning large enough to distinguish frontier models.

Official external benchmark rows are shared across variants of the same model, so a release table attached to a non-reasoning or reasoning tier can still inform the sibling tier instead of creating artificial gaps between variants.

For LiveCodeBench only, the site can synthesize a missing regular value from external LiveCodeBench data. The external row is used as a fallback source rather than as a separate default matrix metric. The fit is learned from overlapping models in the current dataset, then clamped to the normal percentage range before scoring. AIndex does not copy metrics from lower same-family models into higher tiers.

Highly correlated, saturated, or unevenly covered rows such as MMLU-Pro, AIME 2026, and HMMT remain evidence, but they carry lower internal weight so source coverage does not dominate the board.

The weak-prior missing policy is deliberately softer than missing-as-zero, but stricter than ignoring absent boards entirely. It keeps sparse-source models from collapsing while preventing one rich external table from carrying the whole score.

The Custom weights panel exposes the same building blocks for comparison: Best score ratio or Raw score, Geometric Weight Mean or regular Weight Mean, available-only scoring, 0.25 coverage discount, sqrt coverage discount, weak prior, missing-as-zero, full-coverage gates, and per-benchmark weights.