Benchmarks · cross-axis analysis

Axis correlations across the leaderboards

For every pair of leaderboard axes, how much do their per-model rankings agree? Data-derived groupings from pairwise correlation on the models each pair shares, meant to inform how you might weight axes in the composite view — not to define what any axis 'really' measures.

69.7%of the ordering is one dimension

How much of this is one dimension?

Share of the ordering across 47 axes explained by a single underlying factor. A model's position on most boards is largely predictable from its position on the others — which is why a weighted blend mostly re-weights one signal.

One estimate, the conservative one: 276 of 1081 axis pairs share too few models and count as unrelated, and each board's roster restricts the range, so the true share is higher. No imputed or frontier-only variant is shown because none is computed.

General factor loadings

Each axis's share of the shared signal across 47 axes; if every axis counted equally each would be 0.15. Bars are relative to the strongest axis.

Expert0.18
Hard prompts0.18
Overall0.18
Software & IT0.18
Math0.18
English0.18
Instruction following0.18
Coding0.18
Chinese0.18
Business & finance0.18
Russian0.17
French0.17
Multi-turn0.17
Entertainment & media0.17
Writing & language0.17
German0.17
Korean0.17
Japanese0.17
Spanish0.17
Creative writing0.17
GPQA Diamond0.17
AIME (math)0.17
Vision Arena0.16
ARC-AGI-20.16
Factuality Arena0.15
MMLU-Pro (Artificial Analysis)0.15
LiveBench0.15
SimpleBench0.15
FrontierMath0.15
MMLU-Pro (TIGER-Lab)0.14
Humanity's Last Exam0.13
WebDev Arena0.13
LiveCodeBench (Artificial Analysis)0.13
Terminal-Bench0.13
SimpleQA Verified0.13
SWE-bench Verified0.12
BFCL v4 (function calling)0.12
IFBench (Artificial Analysis)0.11
Document Arena0.10
Agent Arena0.08
Text-to-Image Arena0.00
Image Edit Arena0.00
SWE-bench Verified (live board)0.00
Text-to-Video Arena0.00
Search Arena0.00

Left out of the share because they move the other way, so they are measuring something else: Output speed, tokens/s (Artificial Analysis), HHEM Hallucination (Vectara)

The biggest signal

Nearly every axis pair with adequate overlap correlates positively — usually strongly. Distinct 'skill clusters' are weak in this data. Models that are good on one axis tend to be good on many, and 'overall capability' is the dominant shared factor. This does not mean the axes are interchangeable; it means their disagreements are the smaller signal, and their agreement is the larger one.

The extremes, with confidence intervals · cohort ability

The extremes, with confidence intervals

The most and least related pairs, each with a bootstrap interval. These use Spearman rank correlation — the clustering above uses Pearson on the same percentiles. The intervals matter: where a pair shares only twenty-odd models, the interval routinely spans zero, and the coefficient alone would imply a precision the data has not got.

Axis A	Axis B	Spearman rho	95% interval	Shared models
Software & IT	Coding	0.997	0.996 to 0.998	391
Hard prompts	Software & IT	0.997	0.996 to 0.998	396
Overall	English	0.997	0.996 to 0.997	396
Overall	Hard prompts	0.997	0.996 to 0.997	396
Instruction following	Hard prompts	0.997	0.995 to 0.997	396
Hard prompts	English	0.997	0.995 to 0.997	396
Writing & language	Entertainment & media	0.996	0.995 to 0.997	394
Writing & language	Creative writing	0.996	0.994 to 0.996	394
Overall	Instruction following	0.996	0.994 to 0.997	396
Overall	Multi-turn	0.995	0.993 to 0.996	394
Creative writing	Entertainment & media	0.995	0.993 to 0.996	394
English	Software & IT	0.995	0.993 to 0.996	396
Creative writing	Output speed, tokens/s (Artificial Analysis)	-0.03	-0.20 to 0.16spans zero	122
Output speed, tokens/s (Artificial Analysis)	MMLU-Pro (Artificial Analysis)	0.03	-0.12 to 0.19spans zero	159
Output speed, tokens/s (Artificial Analysis)	HHEM Hallucination (Vectara)	-0.02	-0.36 to 0.32spans zero	46
HHEM Hallucination (Vectara)	Terminal-Bench	0.02	-0.44 to 0.49spans zero	23
Writing & language	Output speed, tokens/s (Artificial Analysis)	-0.02	-0.20 to 0.15spans zero	122
Russian	Output speed, tokens/s (Artificial Analysis)	0.02	-0.16 to 0.23spans zero	122
Output speed, tokens/s (Artificial Analysis)	MMLU-Pro (TIGER-Lab)	0.02	-0.30 to 0.35spans zero	39
Entertainment & media	Output speed, tokens/s (Artificial Analysis)	-0.02	-0.21 to 0.17spans zero	122
Math	Output speed, tokens/s (Artificial Analysis)	0.01	-0.20 to 0.20spans zero	116
Overall	Output speed, tokens/s (Artificial Analysis)	-0.00	-0.20 to 0.17spans zero	122
Output speed, tokens/s (Artificial Analysis)	Spanish	0.00	-0.21 to 0.20spans zero	108
Output speed, tokens/s (Artificial Analysis)	French	0.00	-0.20 to 0.22spans zero	103

General capability factor

A single latent ability fitted to every pairwise comparison the boards imply, rather than averaged from percentiles. This is one dimension, not a verdict — it deliberately cannot represent an axis that measures something else.

Fitted only over axes that load positively on the factor above. Excluded: Output speed, tokens/s (Artificial Analysis), HHEM Hallucination (Vectara). Each ability carries the evidence behind it, because a fit without that check put a model measured on a single board at the top.

#	Model	Ability	Boards	Comparisons
1	gemini-4-argon	5.06	22	6,914
2	gpt-5-5-pre-release	4.95	4	506
3	gpt-5-5-pro-pre-release	4.87	3	476
4	claude-opus-5-5	4.79	26	7,469
5	claude-fable-5-1	4.72	30	8,152
6	muse-wonder-max	4.69	17	5,765
7	grok-4-20-0309	4.62	1	323
8	claude-fable-5	4.59	32	8,734
9	claude-opus-5	4.52	31	8,421
10	muse-spark-1-3	4.44	26	7,907
11	grok-4-20-0309-v2	4.42	1	323
12	gemini-3-8-flash	4.39	29	8,299
13	gpt-6-1-sol	4.36	26	7,468
14	claude-opus-4-7	4.31	35	8,878
15	gemini-3-7-flash	4.30	28	8,259

Clusters, singletons and the unclusterable

Correlation is not identity

Two axes that correlate at r ≥ 0.85 (over a sample of at least 20 shared models) share a cluster below, but only in the sense that the rankings agree — not that they measure the same skill. Both might be driven by an underlying variable (model scale, training-data size, training data overlap with the benchmark). Treat cluster membership as 'these rankings agree on the current model population', never as 'these benchmarks measure the same thing'.

Correlation-linked groups (2)

Cluster 1 — 37 axes

Median pairwise r inside the cluster: 0.89

IFBench (Artificial Analysis)
LiveCodeBench (Artificial Analysis)
MMLU-Pro (Artificial Analysis)
BFCL v4 (function calling)
LiveBench
AIME (math)
ARC-AGI-2
FrontierMath
GPQA Diamond
Humanity's Last Exam
SimpleBench
SWE-bench Verified
Terminal-Bench
Chinese
Coding
Creative writing
Document Arena
English
Expert
French
German
Hard prompts
Business & finance
Entertainment & media
Software & IT
Writing & language
Instruction following
Japanese
Korean
Math
Multi-turn
Overall
Russian
Spanish
Factuality Arena
Vision Arena
MMLU-Pro (TIGER-Lab)

Cluster 2 — 2 axes

Median pairwise r inside the cluster: 0.91

Image Edit Arena
Text-to-Image Arena

Axes with adequate overlap but no cluster-worthy correlation (5)

These axes have enough shared models to trust a correlation with peers, but no correlation clears the threshold — their ranking disagrees with everyone else's more than the cutoff allows. That is itself a signal about the axis: it isn't measuring the same thing as the crowd.

Output speed, tokens/s (Artificial Analysis)
SimpleQA Verified
Agent Arena
WebDev Arena
HHEM Hallucination (Vectara)

Axes with insufficient overlap to cluster honestly (9)

These axes have fewer than 20 shared models with every other axis, so any correlation with them would be computed on too small a sample to trust. They are kept in a separate bucket rather than clustered. Once more models are evaluated on them, they may join a cluster in a later run.

GAIA (HAL)
USACO (HAL)
SWE-bench Verified (live board)
τ²-bench Airline
τ²-bench Retail
τ²-bench Telecom
Terminal-Bench 4.0
Search Arena
Text-to-Video Arena

The top 20 pairs

Top 20 pairwise correlations

The raw fact behind the clusters — every pair of axes with adequate overlap, ranked by correlation. Read the clustering step as 'a specific way of grouping this table by a threshold', and if you disagree with the threshold, look here for the underlying numbers.

Axis A	Axis B	Shared models	Pearson r
Coding	Software & IT	391	0.997
Hard prompts	Software & IT	396	0.997
English	Overall	396	0.997
Hard prompts	Overall	396	0.997
Hard prompts	Instruction following	396	0.997
English	Hard prompts	396	0.997
Entertainment & media	Writing & language	394	0.996
Creative writing	Writing & language	394	0.996
Instruction following	Overall	396	0.996
Multi-turn	Overall	394	0.995
Creative writing	Entertainment & media	394	0.995
English	Software & IT	396	0.995
Coding	Hard prompts	391	0.995
English	Instruction following	396	0.994
Software & IT	Overall	396	0.994
Instruction following	Multi-turn	394	0.994
Hard prompts	Multi-turn	394	0.994
Business & finance	Overall	389	0.993
English	Multi-turn	394	0.993
Software & IT	Instruction following	396	0.993