Acropolis

Interactive analysis and open data

Metric explorer

Plot any two benchmark boards against each other across shared models with Spearman, Kendall, and Pearson correlations, or download the open redistribution files directly. Build 17 Nov 2026.

392 shared models
Spearman rho: 0.984Kendall tau-b: 0.899Pearson r: 0.989
1073.7981.61239.81186.01405.91390.4LMArena Text (Arena rating)

Hover or focus any mark to inspect exact values on both benchmarks.

Capability space

Each model as a point in a space of benchmark percentiles, domain composites, principal axes, or domain strengths that remain after the general factor is removed. Choose three axes, drag to rotate, and move the date slider to watch models appear as they were released.

262 models on 33 boards. Missing data rule: a model is plotted only with at least 3 board percentiles. On any axis a model with no value is absent, never filled in.

Drag to rotate, pinch or hold Ctrl and scroll to zoom, arrow keys to turn, brackets to step through points, plus and minus to zoom, 0 to reset.

127 models shown. 135 lack a value on one of these axes. Point size is the number of boards that score the model; faded points have an approximate release date.

X axis, PC1: general factor (66%). Weights: agentic +0.35, coding +0.36, finance +0.32, general-text +0.34, legal +0.26, math +0.31, science +0.39, search +0.3, vision +0.34.

Y axis (up), PC2: +finance -math (20%). Weights: agentic +0.22, coding +0.13, finance +0.49, general-text -0.13, legal +0.49, math -0.46, science -0.08, search -0.44, vision -0.16.

Z axis (depth), PC3: +coding -general text (6%). Weights: agentic +0.48, coding +0.52, finance -0.13, general-text -0.4, legal -0.17, math +0.12, science -0.28, search +0.23, vision -0.39.

Table of the 127 models in view
Models in the current capability space view
ModelLabReleasedWeightsBoardsPC1: general factor (66…PC2: +finance -math (20…PC3: +coding -general t…
GPT-6 AstraOpenAI2026-09-03proprietary253.08-0.831.12
Opus 5.5Anthropic2026-09-22proprietary233.860.70.25
Gemini 3.8 FlashGoogle2026-09-02proprietary231.48-1.78-0.84
Sonnet 5.5Anthropic2026-09-28proprietary223.380.80.92
GPT-6.1 SolOpenAI2026-09-29proprietary222.9-1.450.85
GPT-6 LunaOpenAI2026-09-22proprietary21-1.35-1.40.23
Grok 4.7xAI2026-09-21proprietary210.661.020.36
Haiku 5.5Anthropic2026-10-07proprietary200.16-0.250.15
GLM-5.3 FlashZ.ai2026-08-20open200.47-1.75-0.85
Gemini 4 ArgonGoogle2026-10-13proprietary193.720.740.03
Opus 4.6Anthropic2026-02-05proprietary183.530.420.68
Gemini 4 CarbonGoogle2026-11-10proprietary184.30.750.46
Kimi K3Moonshot AI2026-07-16open181.22-0.69-0.53
GLM-5.3Z.ai2026-08-14open181.8-0.18-0.5
Opus 4.7Anthropic2026-04-16proprietary173.210.470.46
Sonnet 4.6Anthropic2026-02-17proprietary171.38-0.090.23
Gemini 3 Pro PreviewGoogle2025-11-18proprietary171.3-0.280.07
Gemini 3.1 Pro PreviewGoogle2026-02-19proprietary172.01-0.55-0.11
GPT-5.2OpenAI2025-12-11proprietary170.710.830.64
GPT-5.4OpenAI2026-03-05proprietary172.270.690.32
Inkling SmallThinking Machines2026-07-30open17-2.69-2.17-0.81
Fable 5.1Anthropic2026-09-01proprietary163.75-0.020.49
Opus 4.5Anthropic2025-11-24proprietary160.891.770.7
Opus 5Anthropic2026-07-24proprietary163.11-0.490.23
Muse Wonder 1.0Meta2026-10-12proprietary162.70.38-0.16
GPT-5.1OpenAI2025-11-13proprietary16-1.39-0.64-0.21
GPT-5.6 SolOpenAI2026-07-09proprietary162.86-0.980.55
InklingThinking Machines2026-07-15open16-1.73-2.63-1.14
MiMo-V2.6-FlashXiaomi2026-09-21unknown16-0.03-1.28-0.47
MiMo-V2.6-ProXiaomi2026-09-21unknown161.03-0.35-0.4
Fable 5Anthropic2026-06-09proprietary153.11-0.720.1
Fable 5.2Anthropic2026-11-09proprietary154.75-0.270.66
Opus 4.8Anthropic2026-05-28proprietary151.60.19-0.16
Sonnet 4.5Anthropic2025-09-29proprietary15-1.761.040.37
Sonnet 5Anthropic2026-06-30proprietary150.550.660.04
DeepSeek V4 Pro 0813DeepSeek2026-08-13open150.11-1.58-0.64
DeepSeek V4.1 Flash 1017DeepSeek2026-10-17open151.92-0.150.13
Gemini 2.5 ProGoogle2025-06-05proprietary15-3.1-0.23-0.49
Gemini 3 Flash PreviewGoogle2025-12-17proprietary150.371.030.43
Gemini 4 HydrogenGoogle2026-11-12proprietary153.470.60.38
Kimi K2.5Moonshot AI2026-01-27open15-0.30.230.18
Kimi K3.1 ProMoonshot AI2026-10-23open152.35-1.350.13
Step 5 PreviewStepFun2026-09-18unknown15-0.22-1.14-0.58
GPT-5OpenAI2025-08-07proprietary14-0.161.090.34
GPT-5.5OpenAI2026-04-23proprietary142.3-0.850.29
GPT-5.6 TerraOpenAI2026-07-09proprietary141.01-1.13-0.2
GPT-6.1 TerraOpenAI2026-10-26proprietary145.48-4.28-0.51
Beam 1.0Reflection2026-10-16open14-1.22-2.52-1.01
Inkling 2 SmallThinking Machines2026-11-05unknown14-0.361.10.07
Grok 4.5xAI2026-07-08proprietary140.31-0.830.22
Grok 4.6xAI2026-08-12proprietary140.82-0.180.11
GPT-5.6 LunaOpenAI2026-07-09proprietary13-0.17-0.070.07
GPT-6 SolOpenAI2026-09-22proprietary132.430.910.71
o3OpenAI2025-04-16proprietary13-1.77-0.28-0.24
Opus 4Anthropic2025-05-22proprietary12-2.261.60.12
Gemini 3.5 FlashGoogle2026-05-19proprietary122.050.420.42
Gemini 3.7 FlashGoogle2026-08-13proprietary121.52-2.02-0.39
Muse Spark OpenMeta2026-10-15open12-1.6-1.16-0.7
Muse Wonder Full 1.0Meta2026-10-12unknown121.941.670.78
Hy4 PreviewTencent2026-08-28 (approx.)unknown12-0.81-0.76-0.44
Claude 3.7 SonnetAnthropic2025-02-24proprietary11-2.942.20.28
Haiku 4.5Anthropic2025-10-15proprietary11-4.051.260.04
Opus 4.1Anthropic2025-08-05proprietary11-1.091.19-0.26
Muse EmberMeta2026-11-09unknown110.121.720.57
MiniMax M3MiniMax2026-06-01open11-2.44-1.1-0.58
MiniMax M3.1 FlashMiniMax2026-10-19unknown11-0.38-1.41-0.41
Kimi K2.6Moonshot AI2026-04-20open110.990.880.41
GPT-5 miniOpenAI2025-08-07proprietary11-2.26-0.16-0.1
Qwen3.8-MaxAlibaba2026-08-02proprietary101.92-2.58-0.49
Gemini 3.6 FlashGoogle2026-07-21proprietary100.44-0.72-0.07
Muse Spark 1.3Meta2026-09-02unknown102.08-0.790.04
Mistral Large 4 OSSMistral2026-10-29open100-0.67-0.13
Hy4Tencent2026-11-03unknown101.46-0.370.13
Qwen3.7 PlusAlibaba2026-06-01proprietary9-1.83-2.19-0.83
Qwen3.8 Max InternalAlibaba2026-11-15 (approx.)unknown90.410.650.26
Qwen4 MaxAlibaba2026-11-15 (approx.)proprietary91.490.880.47
Claude Sonnet 4Anthropic2025-05-26 (approx.)proprietary9-1.952.60.41
DeepSeek V4.1 FlashDeepSeek2026-09-09open91.45-0.56-0.02
Gemma 4 31BGoogle2026-04-02open9-1.48-2.16-0.77
Mercury 3Inception2026-11-15 (approx.)unknown9-1.82-2.02-0.8
Muse SparkMeta2026-04-08proprietary91.37-1.31-0.35
Muse Spark 1.1Meta2026-07-09proprietary90.95-1.84-0.46
Mistral Large 4Mistral2026-10-06open9-1.2-0.72-0.34
Nemotron 3.5 Super VLNVIDIA2026-10-15open9-2.29-2.3-0.96
GPT-4.1OpenAI2025-04-14proprietary9-2.692.680.42
GPT-5.4 miniOpenAI2026-03-17proprietary9-1.13-0.71-0.37
Grok 4.3xAI2026-04-30proprietary9-0.270.730.14
GLM-5Z.ai2026-02-11open9-0.19-0.32-0.25
GLM-5.2Z.ai2026-06-16open91.4-0.120.03
Qwen3.8-27BAlibaba2026-08-14open8-0.74-1.79-0.61
Claude 3.5 Sonnet (Oct '24)Anthropic2024-10-22proprietary8-2.353.810.76
DeepSeek V4-ProDeepSeek2026-04-24open80.48-2.87-0.8
GPT-5 nanoOpenAI2025-08-07proprietary8-2.32.570.51
o1OpenAI2024-12-05proprietary8-1.362.520.55
o4-miniOpenAI2025-04-16proprietary8-1.920.560.12
Grok 4 (0709)xAI2025-07-09proprietary8-0.671.550.39
Grok 4.1 FastxAI2025-11-19proprietary8-0.2910.53
Qwen3.6 Max PreviewAlibaba2026-04-20unknown71.62-0.76-0.04
Qwen3.6 PlusAlibaba2026-04-02proprietary7-0.09-2.02-0.64
DeepSeek V3.2DeepSeek2025-12-01open7-1.120.22-0.15
Gemini 1.5 Pro (002)Google2024-09-24proprietary7-3.52.880.45
Gemini 2.0 Flash (001)Google2025-02-05proprietary7-1.862.670.57
Gemini 2.5 FlashGoogle2025-05-20proprietary7-2.13-1.45-0.7
Gemini 3.1 Flash-Lite PreviewGoogle2026-03-03proprietary7-0.271.20.29
Gemini 3.5 Flash-LiteGoogle2026-07-21proprietary70.510.360.11
Llama 4 ScoutMeta2025-04-05open7-3.133.310.63
Muse Spark 1.2Meta2026-08-05unknown70.68-4.01-1.15
Mistral Medium 3Mistral2025-05-07proprietary7-1.024.641.01
GPT-4.1 miniOpenAI2025-04-14proprietary7-1.952.490.49
GPT-4.1 nanoOpenAI2025-04-14proprietary7-3.093.770.77
GPT-4.5 PreviewOpenAI2025-02-28 (approx.)proprietary7-2.412.380.39
GPT-5.4 nanoOpenAI2026-03-17proprietary7-1.291.810.39
GLM-4.7Z.ai2025-12-22open70.05-1.76-0.52
GLM-5.1Z.ai2026-04-07open71.61-1.25-0.21
Llama 4 MaverickMeta2025-04-05open6-3.633.280.55
MiniMax M2.5MiniMax2026-02-12open6-1.850.4-0.03
Mistral Medium 3.5Mistral2026-04-29unknown6-2.2-1.47-0.75
Nemotron 3.5 Super VL 120B A12B (NVFP4)NVIDIA2026-11-09 (approx.)open6-1.95-0.71-0.47
GPT-4o miniOpenAI2024-07-18proprietary6-3.163.950.8
gpt-oss-120bOpenAI2025-08-05open6-3.1-1.2-0.7
Grok 4.20 Multi-Agent Beta (0309)xAI2026-11-09 (approx.)proprietary61.26-0.060.14
MiniMax M2MiniMax2025-10-26open5-2.740.21-0.23
GPT-4o (Nov '24)OpenAI2024-11-20proprietary5-5.683.580.58
gpt-oss-20bOpenAI2025-08-05open5-4.32-0.05-0.49
DeepSeek V4-Flash 0731DeepSeek2026-07-31open4-1.23-2.58-0.89
Gemini 2.5 Pro Preview (06-05)Google2025-06-05proprietary4-3.01-0.04-0.2
GPT-5.4 ProOpenAI2026-03-05proprietary41.1-1.66-0.37
Method, limits, and excluded inputs (33 boards, 2660 rows read)

Method rules

  • Boards: one default board per benchmark and source, quality metrics only, best visible row per model. A board needs at least 12 scored models to be an axis.
  • Board axis: the model's percentile inside that board, 0 (worst) to 100 (best), ties sharing the midpoint.
  • Domain composite: the median percentile over the boards of the domain, where a board counts toward every domain its benchmark lists. A domain needs 2 boards, and a model needs 2 of them. Fewer means no value, never a zero.
  • Missing data rule: a model is plotted only with at least 3 board percentiles. On any axis a model with no value is absent, never filled in.
  • Principal axes: PCA of the standardized domain composites (domains with at least 25 models; models with at least 3 of them). Loadings come from the pairwise-complete correlations of the domains. A gap is filled by rank-2 iterative imputation (30 rounds) only to project the model onto them. PC1 is signed so a higher value means stronger overall; the other signs are set so the largest loading is positive.
  • Specialization: for each domain, the model's standardized composite minus the part the general factor (PC1) predicts. Zero means the domain matches the model's overall level; positive means stronger than overall. Only observed domains get a value.
  • Release date comes from the registry; where it is missing, the earliest observation date of the model's rows is used and the model is marked as approximate.

Known limits

  • Percentiles are relative to the models each board happens to test, so a board of frontier models and a board of small models are not on one scale.
  • Boards that share a domain often share models and tasks, so a domain composite is less independent than its board count suggests.
  • Principal axes are fitted to this set of models and boards and change when either changes. They describe correlation, not a causal skill.
  • A model with few observed domains is placed on the principal axes partly by imputation, so its position is less certain.

Exclusions

  • 17 boards excluded: board has fewer than 12 scored models
  • 94 boards excluded: not the default slice of its source and version
  • 1 boards excluded: older version of a benchmark the source reports again
  • 2 boards excluded: rank metric is not a quality score
  • 135 models excluded: model has fewer than 3 observed domains (no principal axes)
  • 372 models excluded: model scored on fewer than 3 qualifying boards
  • 75 rows excluded: hidden row (for example superseded)

Open bulk exports

Static CSV, JSONL, and search-index artifacts emitted during build. Restricted sources (Artificial Analysis) are automatically excluded from bulk downloads.

DuckDB quick query

SELECT benchmarkId, modelId, value, ciLow, ciHigh
FROM read_csv_auto('acropolis-open-results.csv')
WHERE benchmarkId = 'swe-bench-verified'
ORDER BY value DESC
LIMIT 15;

Python Polars quick query

import polars as pl
df = pl.read_ndjson("acropolis-open-results.jsonl")
print(df.group_by("benchmarkId").len().sort("len", descending=True))