Frontier model ranking based on top-tier private benchmarks · Higher is better
Top per category vs. lowest · Higher is better
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 | Knowledge WorkBusiness | 10S | 10 | 10 | 10 | 10 | ||
| 2 | Knowledge WorkBusiness | 10S | 10 | 10 | 10 | 10 | ||
| 3 | ScienceMath, Science | 10S | 10 | 10 | 10 | 10 | ||
| 4 | Knowledge WorkLegal | 10S | 10 | 10 | 10 | 10 | ||
| 5 | ScienceMedical | 10S | 10 | 10 | 10 | 10 | ||
| 6 | CodingCoding | 9.6S | 10 | 10 | 10 | 7.3 | ||
| 7 | CodingCoding | 9.3S | 10 | 7.2 | 10 | 10 | ||
| 8 | Knowledge WorkBusiness, Finance +3 | 9S | 10 | 10 | 10 | 3.3 | ||
| 9 | Knowledge WorkBusiness, Finance | 9S | 10 | 10 | 10 | 3.3 | ||
| 10 | CodingCoding | 8.8S | 7.5 | 10 | 9 | 10 | ||
| 11 | Knowledge WorkBusiness | 8.8S | 7.5 | 10 | 9 | 10 | ||
| 12 | CodingCoding, Cyber | 8.8S | 7.5 | 10 | 9 | 10 | ||
| 13 | Knowledge WorkFinance, Business +1 | 8.3A | 10 | 7.2 | 10 | 3.3 | ||
| 14 | GeneralHumanities, Business +4 | 8.3A | 7.5 | 7.2 | 10 | 10 | ||
| 15 | GeneralMath, Coding +2 | 8.3A | 7.5 | 7.2 | 10 | 10 | ||
| 16 | GeneralMath, Science | 8.2A | 7.5 | 10 | 6 | 10 | ||
| 17 | ScienceScience, Coding | 8A | 5.5 | 10 | 9 | 10 | ||
| 18 | CodingScience, Humanities +1 | 7.8A | 5.5 | 10 | 8 | 10 | ||
| 19 | CodingCoding | 7.5A | 10 | 0 | 10 | 10 | ||
| 20 | CodingCyber, Coding | 7.5A | 10 | 0 | 10 | 10 | ||
| 21 | CodingCoding | 7.5A | 7.5 | 4 | 10 | 10 | ||
| 22 | Knowledge WorkBusiness, Finance +4 | 7A | 5.5 | 10 | 9 | 3.3 | ||
| 23 | GeneralScience, Math +2 | 5.8B | 2 | 10 | 7 | 7.3 | ||
| 24 | GeneralBusiness, Science +1 | 5.6B | 5.5 | 10 | 2 | 3.3 | ||
| 25 | Knowledge WorkCoding, Business | 5.5B | 2 | 10 | 3.5 | 10 | ||
| 26 | ScienceScience | 5.4B | 0 | 10 | 7 | 10 | ||
| 27 | Knowledge WorkBusiness, Coding +2 | 5.2B | 2 | 10 | 7 | 3.3 | ||
| 28 | ScienceMath | 4.7C | 0 | 7.2 | 7 | 10 | ||
| 29 | CodingCoding | 4.7C | 0 | 7.2 | 7 | 10 | ||
| 30 | ScienceScience | 4.5C | 2 | 7.2 | 7 | 3.3 | ||
| 31 | ScienceScience | 4.4C | 2 | 10 | 0 | 7.3 | ||
| 32 | CodingCoding, Cyber +1 | 4.4C | 0 | 10 | 7 | 3.3 | ||
| 33 | CodingCoding, Cyber +1 | 4.4C | 0 | 10 | 7 | 3.3 | ||
| 34 | CodingCoding | 4.3C | 0 | 7.2 | 7 | 7.3 | ||
| 35 | ScienceMedical | 4.3C | 0 | 10 | 3.5 | 7.3 | ||
| 36 | ScienceScience, Coding | 4.2C | 0 | 10 | 6 | 3.3 | ||
| 37 | Knowledge WorkScience, Finance +2 | 4.1C | 0 | 7.2 | 6 | 7.3 | ||
| 38 | CodingCoding | 3.9C | 0 | 4 | 7 | 10 | ||
| 39 | CodingCoding | 3.7C | 0 | 4 | 6 | 10 | ||
| 40 | Knowledge WorkBusiness, Finance +3 | 3.6C | 0 | 7.2 | 3.5 | 7.3 | ||
| 41 | Knowledge WorkBusiness, Finance +3 | 3.3C | 0 | 10 | 1.5 | 3.3 | ||
| 42 | GeneralCoding, Humanities | 3.2C | 0 | 7.2 | 1.5 | 7.3 | ||
| 43 | CodingCoding, Math | 3.2C | 0 | 7.2 | 1.5 | 7.3 | ||
| 44 | GeneralScience, Humanities +3 | 2.1D | 0 | 4 | 0 | 7.3 | ||
| 45 | CodingCoding | 1.5D | 0 | 4 | 0 | 3.3 | ||
| 46 | GeneralScience, Math +4 | 1.1D | 0 | 0 | 0 | 7.3 | ||
| 47 | CodingCyber, Coding | 0.5D | 0 | 0 | 0 | 3.3 |
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.
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.
| # | Model | Top Tier (S/A) | Low Tier (C/D) | Low-Tier Boost |
|---|---|---|---|---|
| 1 | GPT-6.2 Astra max | 96.9p | 85.5p | -2.5p |
| 2 | Claude Fable 5.2 max | 95.3p | 88.5p | -6.6p |
| 3 | Gemini 4 Neon xhigh | 93.1p | 91.5p | +0.2p |
| 4 | Gemini 4 Carbon high | 84.6p | 87.4p | +3.9p |
| 5 | GPT-6.2 Sol max | 83.8p | 78.4p | -0.7p |
| 6 | Claude Opus 5.5 max | 83.0p | 86.6p | +11.2p |
| 7 | GPT-6 Astra max | 83.0p | 77.3p | -0.4p |
| 8 | 78.3p | 82.8p | +1.6p | |
| 9 | Muse Wonder Full max | 78.0p | 81.6p | +1.1p |
| 10 | GPT-6.1 Sol max | 77.7p | 76.4p | +1.9p |
| 11 | Gemini 4 Hydrogen xhigh | 76.7p | 74.1p | -0.4p |
| 12 | Claude Fable 5.1 max | 70.7p | 78.9p | +14.2p |
| 13 | GPT-6.1 Terra max | 70.5p | 65.3p | -5.3p |
| 14 | Muse Wonder max | 68.1p | 68.9p | -7.4p |
| 15 | Gemini 4 Argon high | 67.9p | 66.4p | +5.2p |
| 16 | GPT-6 Sol max | 66.4p | 58.0p | -6.6p |
| 17 | Kimi K3.1 Pro max | 65.3p | 82.1p | +20.3p |
| 18 | Qwen4 Max xhigh | 61.5p | 69.5p | +8.8p |
| 19 | 56.0p | 48.9p | -6.4p | |
| 20 | Claude Haiku 5.5 max | 55.3p | 51.1p | -12.7p |
| 21 | Grok 4.7 xhigh | 53.3p | 49.8p | -2.7p |
| 22 | Tencent Hy4 default | 52.1p | 53.1p | -4.4p |
| 23 | Muse Ember max | 50.8p | 59.0p | +4.9p |
| 24 | Inkling 2 Small default | 49.6p | 53.8p | +5.1p |
| 25 | Muse Spark 1.3 max | 49.6p | 55.8p | +16.2p |
| 26 | 48.3p | 39.0p | -6.9p | |
| 27 | Qwen3.8 Max (0902) xhigh | 43.8p | 55.7p | +14.0p |
| 28 | GPT-6 Luna max | 40.9p | 34.1p | -10.5p |
| 29 | GLM-5.3 max | 40.8p | 51.8p | +6.0p |
| 30 | MiMo-V2.6-Pro default | 40.7p | 48.1p | +4.7p |
| 31 | Gemini 3.7 Flash high | 39.4p | 42.9p | +19.2p |
| 32 | Ling 3.1 Flash 1019 default | 37.3p | 36.9p | -5.0p |
| 33 | 36.2p | 34.2p | +0.3p | |
| 34 | Qwen3.8-Flash-Next default | 34.6p | 32.3p | +1.0p |
| 35 | Gemini 3.8 Flash high | 32.6p | 42.2p | +14.9p |
| 36 | Step 5 Preview default | 31.8p | 30.0p | -10.3p |
| 37 | MiMo-V2.6-Flash default | 27.8p | 23.5p | -3.6p |
| 38 | Mistral Large 4 high | 24.7p | 32.8p | +10.4p |
| 39 | Qwen3.8 27B xhigh | 23.7p | 13.3p | -10.1p |
| 40 | Muse Spark Open xhigh | 20.8p | 27.1p | +6.5p |
| 41 | Beam 1.0 default | 18.3p | 10.1p | -3.7p |
| 42 | K2 Horizon 375B A23B high | 11.8p | 4.4p | -4.7p |
| 43 | 10.4p | 6.1p | -1.2p | |
| 44 | Apodex 1.1 default | 2.3p | 11.9p | +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.