Frontier model ranking based on top-tier private benchmarks · Higher is better

05010096.995.393.184.683.883.083.078.378.077.776.770.7

Top per category vs. lowest · Higher is better

Lowest audit score

Showing 47 audited benchmarks

Score & factor definitions (BenchAudit Score, Privacy, Originality, Freshness, Grading)
BenchAudit Score
Weighted composite of Privacy, Originality, Freshness, and Grading on a 0–10 scale. Higher is better — stronger resistance to contamination and grading leaks.
Privacy
How strictly test prompts and answers are kept private versus open to public downloads, GitHub repos, or auto-approved click-through gates.
Originality
How much of the benchmark was newly authored from scratch versus reused from older public exams, competition archives, or web corpora.
Freshness
How recently tasks were refreshed relative to exposure risk, including rolling post-cutoff updates and leaked-task pruning.
Grading
Whether models are scored by an independent evaluator inside an isolated offline sandbox, free from live-internet or container/git leaks.
# Benchmark & Hub Category & Domains Audit
1
aa-briefcase-v1-1· Artificial Analysis
Knowledge WorkBusiness
10S10101010
2
automationbench-aa· Artificial Analysis / Zapier
Knowledge WorkBusiness
10S10101010
3
frontiermath· Epoch AI
ScienceMath, Science
10S10101010
4
harvey-lab-aa-v1-1· Artificial Analysis / Harvey
Knowledge WorkLegal
10S10101010
5
mlcr-aa· Artificial Analysis / Wisedocs
ScienceMedical
10S10101010
6
vcb-1-100-v1-1· Vals AI
CodingCoding
9.6S1010107.3
7
frontiercode-aa-2-0· Artificial Analysis / Cognition
CodingCoding
9.3S107.21010
8
mka· Artificial Analysis / Google DeepMind
Knowledge WorkBusiness, Finance +3
9S1010103.3
9
vending-bench-2· Andon Labs
Knowledge WorkBusiness, Finance
9S1010103.3
10
code2video-aa· Artificial Analysis / HeyGen
CodingCoding
8.8S7.510910
11
enterpriseops-gym-aa· Artificial Analysis / ServiceNow
Knowledge WorkBusiness
8.8S7.510910
12
itbench-aa· Artificial Analysis / IBM
CodingCoding, Cyber
8.8S7.510910
13
aa-analystagent· Artificial Analysis
Knowledge WorkFinance, Business +1
8.3A107.2103.3
14
aa-omniscience-v1-1· Artificial Analysis
GeneralHumanities, Business +4
8.3A7.57.21010
15
livebench· LiveBench / Abacus.AI
GeneralMath, Coding +2
8.3A7.57.21010
16
arc-agi-2· ARC Prize Foundation
GeneralMath, Science
8.2A7.510610
17
critpt-v2· Artificial Analysis / CritPt
ScienceScience, Coding
8A5.510910
18
aa-3d· Artificial Analysis
CodingScience, Humanities +1
7.8A5.510810
19
code-migration-v1-1· Vals AI
CodingCoding
7.5A1001010
20
cyberbench· Vals AI
CodingCyber, Coding
7.5A1001010
21
swe-bench-pro· Scale AI
CodingCoding
7.5A7.541010
22
agents-last-exam-v1-1· UC Berkeley RDI
Knowledge WorkBusiness, Finance +4
7A5.51093.3
23
hle-diamond· Artificial Analysis / CAIS / Scale AI
GeneralScience, Math +2
5.8B21077.3
24
gaia· Hugging Face / Meta / AutoGPT
GeneralBusiness, Science +1
5.6B5.51023.3
25
apex-agents-aa· Artificial Analysis / Mercor
Knowledge WorkCoding, Business
5.5B2103.510
26
vials-aa· Artificial Analysis / Handshake AI
ScienceScience
5.4B010710
27
osworld-2-1-offline· OSWorld / XLANG Lab (HKU)
Knowledge WorkBusiness, Coding +2
5.2B21073.3
28
autoform-hard· Vals AI
ScienceMath
4.7C07.2710
29
swe-atlas-qna· Scale AI
CodingCoding
4.7C07.2710
30
biomysterybench· Vals AI
ScienceScience
4.5C27.273.3
31
gpqa-diamond· Artificial Analysis / NYU
ScienceScience
4.4C21007.3
32
terminal-bench-4-0· Artificial Analysis / Harbor
CodingCoding, Cyber +1
4.4C01073.3
33
terminal-bench-4-1· Artificial Analysis / Harbor
CodingCoding, Cyber +1
4.4C01073.3
34
deepswe· Datacurve
CodingCoding
4.3C07.277.3
35
healthbench· Artificial Analysis / OpenAI
ScienceMedical
4.3C0103.57.3
36
terminal-bench-science-0-1· Artificial Analysis / Stanford / Harbor
ScienceScience, Coding
4.2C01063.3
37
chartography· Surge AI
Knowledge WorkScience, Finance +2
4.1C07.267.3
38
swe-bench-pro-v2· Scale AI / Reflection
CodingCoding
3.9C04710
39
swe-sweep· Meta Superintelligence Labs
CodingCoding
3.7C04610
40
gdp-pdf· Artificial Analysis / Surge AI
Knowledge WorkBusiness, Finance +3
3.6C07.23.57.3
41
gdpval-aa-v2-1· Artificial Analysis / OpenAI
Knowledge WorkBusiness, Finance +3
3.3C0101.53.3
42
ifbench· Artificial Analysis / AllenAI
GeneralCoding, Humanities
3.2C07.21.57.3
43
livecodebench· LiveCodeBench / UC Berkeley
CodingCoding, Math
3.2C07.21.57.3
44
mmmu-pro· Artificial Analysis / MMMU
GeneralScience, Humanities +3
2.1D0407.3
45
swe-bench-verified· SWE-bench / Princeton / OpenAI
CodingCoding
1.5D0403.3
46
mmlu-pro· TIGER-AI-Lab / Artificial Analysis
GeneralScience, Math +4
1.1D0007.3
47
cybench· Cybench / Stanford
CodingCyber, Coding
0.5D0003.3

Model Integrity: Real Capability vs Benchmark Gaming

See which models hold up on private, leak-resistant evaluations, and which ones collapse once public training data leaks are removed.

44 frontier models · 38 evals
Min. Tests 7+ matched · 3+ domains

How to read the True Index: Rankings are built exclusively from Top-Tier Tests (Tiers S & A)—private, recently refreshed evaluations where memorized answers can’t inflate scores. Each model’s percentile is balanced equally across Knowledge Work, Coding, Science, and General (25% each, renormalized across covered categories) to surface genuine capability rather than leaderboard polish.

#ModelTop Tier (S/A)Low Tier (C/D)Low-Tier Boost
1
OpenAI· 21 matched evals· 4 domains
96.9p (13)85.5p (8)-2.5p
2
Anthropic· 18 matched evals· 3 domains
95.3p (13)88.5p (8)-6.6p
3
Google· 8 matched evals· 3 domains
93.1p (7)91.5p (3)+0.2p
4
Google· 12 matched evals· 3 domains
84.6p (10)87.4p (5)+3.9p
5
OpenAI· 15 matched evals· 3 domains
83.8p (11)78.4p (7)-0.7p
6
Anthropic· 25 matched evals· 4 domains
83.0p (17)86.6p (8)+11.2p
7
OpenAI· 24 matched evals· 4 domains
83.0p (16)77.3p (8)-0.4p
8
Anthropic· 21 matched evals· 3 domains
78.3p (15)82.8p (8)+1.6p
9
Meta· 13 matched evals· 3 domains
78.0p (9)81.6p (5)+1.1p
10
OpenAI· 25 matched evals· 4 domains
77.7p (16)76.4p (9)+1.9p
11
Google· 9 matched evals· 3 domains
76.7p (8)74.1p (4)-0.4p
12
Anthropic· 17 matched evals· 3 domains
70.7p (14)78.9p (6)+14.2p
13
OpenAI· 15 matched evals· 4 domains
70.5p (9)65.3p (6)-5.3p
14
Meta· 17 matched evals· 3 domains
68.1p (12)68.9p (6)-7.4p
15
Google· 21 matched evals· 3 domains
67.9p (16)66.4p (8)+5.2p
16
OpenAI· 18 matched evals· 4 domains
66.4p (11)58.0p (7)-6.6p
17
Kimi· 17 matched evals· 4 domains · Open Weights
65.3p (11)82.1p (6)+20.3p
18
Qwen4 Max xhigh
Alibaba· 10 matched evals· 3 domains
61.5p (8)69.5p (4)+8.8p
19
DeepSeek· 19 matched evals· 4 domains · Open Weights
56.0p (13)48.9p (6)-6.4p
20
Anthropic· 21 matched evals· 3 domains
55.3p (15)51.1p (8)-12.7p
21
Grok 4.7 xhigh
SpaceXAI· 20 matched evals· 3 domains
53.3p (17)49.8p (6)-2.7p
22
Tencent Hy4 default
Tencent· 10 matched evals· 3 domains · Open Weights
52.1p (7)53.1p (4)-4.4p
23
Meta· 7 matched evals· 3 domains
50.8p (5)59.0p (3)+4.9p
24
Thinking Machines· 18 matched evals· 4 domains · Open Weights
49.6p (10)53.8p (8)+5.1p
25
Meta· 15 matched evals· 3 domains
49.6p (13)55.8p (4)+16.2p
26
DeepSeek· 12 matched evals· 3 domains · Open Weights
48.3p (7)39.0p (6)-6.9p
27
Alibaba· 18 matched evals· 4 domains
43.8p (13)55.7p (5)+14.0p
28
OpenAI· 24 matched evals· 4 domains
40.9p (16)34.1p (8)-10.5p
29
Z AI· 18 matched evals· 3 domains · Open Weights
40.8p (14)51.8p (5)+6.0p
30
Xiaomi· 16 matched evals· 3 domains · Open Weights
40.7p (14)48.1p (4)+4.7p
31
Google· 9 matched evals· 3 domains
39.4p (8)42.9p (4)+19.2p
32
InclusionAI· 9 matched evals· 3 domains · Open Weights
37.3p (6)36.9p (4)-5.0p
33
MiniMax· 16 matched evals· 4 domains · Open Weights
36.2p (9)34.2p (7)+0.3p
34
Alibaba· 8 matched evals· 3 domains · Open Weights
34.6p (6)32.3p (4)+1.0p
35
Google· 23 matched evals· 4 domains
32.6p (16)42.2p (7)+14.9p
36
StepFun· 18 matched evals· 4 domains · Open Weights
31.8p (12)30.0p (6)-10.3p
37
Xiaomi· 16 matched evals· 4 domains · Open Weights
27.8p (10)23.5p (6)-3.6p
38
Mistral· 10 matched evals· 3 domains · Open Weights
24.7p (7)32.8p (4)+10.4p
39
Alibaba· 12 matched evals· 3 domains · Open Weights
23.7p (10)13.3p (4)-10.1p
40
Meta· 16 matched evals· 4 domains · Open Weights
20.8p (10)27.1p (6)+6.5p
41
Beam 1.0 default
Reflection· 13 matched evals· 3 domains · Open Weights
18.3p (8)10.1p (6)-3.7p
42
Institute of Foundation Models· 9 matched evals· 3 domains · Open Weights
11.8p (6)4.4p (4)-4.7p
43
NVIDIA· 11 matched evals· 3 domains · Open Weights
10.4p (7)6.1p (5)-1.2p
44
Apodex 1.1 default
Apodex· 10 matched evals· 4 domains
2.3p (5)11.9p (5)+8.5p

BenchAudit is an independent index scoring AI benchmarks on resistance to training-data contamination and grading leaks, paired with the BenchAudit True Index showing how frontier models hold up on clean, private evaluations. Explore the CC BY 4.0 dataset on Hugging Face.