Benchmarks · Safety & security
HoneyBench v0.2
Does the model cheat when it can? Tasks with clear rules and a flawed scorer, such as dosing simulated patients, grid control and ML repair. The share of scored runs whose final submission exploits the flaw. Lower is better.
22 models · 22 results · data as of 2026-11-12 · goodhartlabs.com ↗
Leaderboard
Fable 5.2 max
More from this benchmark
54 published measuresScore vs any published measure54 measures: sub-scores, splits, counts
11 more models under All models in settings.
Breakdown
darker is better within each column · click a column to sort| Model | |||
|---|---|---|---|
| Fable 5.2 max | 80.0 | 8 | 10.0 |
| Opus 5.5 max | 80.0 | 8 | 10.0 |
| Sonnet 5.5 max | 100.0 | 10.0 | 10.0 |
| GPT-6 Astra max | 100.0 | 10.0 | 10.0 |
| Inkling 2 Small max | 100.0 | 10.0 | 10.0 |
| GPT-6.1 Sol max | 80.0 | 8 | 10.0 |
| Fable 5.1 max | 60.0 | 6 | 10.0 |
| Gemini 4 Argon max | 100.0 | 9 | 9 |
| Step 5 Preview high | 90.0 | 9 | 10.0 |
| Qwen 3.8 Max | 100.0 | 10.0 | 10.0 |
| DeepSeek V4 Pro max | 100.0 | 10.0 | 10.0 |
| Kimi K3.1 max | 100.0 | 10.0 | 10.0 |
| DeepSeek V4.1 Flash 1017 max | 100.0 | 10.0 | 10.0 |
| GLM 5.3 max | 100.0 | 10.0 | 10.0 |
| Kimi K3 max | 100.0 | 10.0 | 10.0 |
| Gemini 3.8 Flash max | 100.0 | 10.0 | 10.0 |
| Muse Wonder max | 100.0 | 10.0 | 10.0 |
| DeepSeek V4.1 Flash max | 100.0 | 10.0 | 10.0 |
| Mistral Large 4 max | 100.0 | 10.0 | 10.0 |
| Muse Spark 1.3 max | 100.0 | 10.0 | 10.0 |
| Grok 4.7 max | 100.0 | 10.0 | 10.0 |
| Muse Spark Open xhigh | 100.0 | 10.0 | 10.0 |
All results
22 rows| # | Model | Effort | HoneyBench v0.2 |
|---|---|---|---|
| 1 | Fable 5.2 | max | 11.9% ±3.2 |
| 2 | Opus 5.5 | max | 31.6% ±4.8 |
| 3 | Sonnet 5.5 | max | 41.6% ±4.6 |
| 4 | GPT-6 Astra | max | 50.8% ±4.5 |
| 5 | Inkling 2 Small | max | 52.6% ±4.9 |
| 6 | GPT-6.1 Sol | max | 59.5% ±4.5 |
| 7 | Fable 5.1 | max | 61.8% ±4.8 |
| 8 | Gemini 4 Argon | max | 63.6% ±5.3 |
| 9 | Step 5 Preview | high | 64.9% ±4.5 |
| 10 | Qwen 3.8 Max | max | 66.2% ±4.3 |
| 11 | DeepSeek V4 Pro | max | 71.4% ±4.3 |
| 12 | Kimi K3.1 | max | 71.8% ±4.1 |
| 13 | DeepSeek V4.1 Flash 1017 | max | 76.4% ±3.9 |
| 14 | GLM 5.3 | max | 76.8% ±4.0 |
| 15 | Kimi K3 | max | 78.5% ±3.8 |
| 16 | Gemini 3.8 Flash | max | 80.5% ±3.7 |
| 17 | Muse Wonder | max | 81.0% ±3.6 |
| 18 | DeepSeek V4.1 Flash | max | 81.0% ±3.6 |
| 19 | Mistral Large 4 | max | 83.6% ±3.4 |
| 20 | Muse Spark 1.3 | max | 83.6% ±3.4 |
| 21 | Grok 4.7 | max | 85.1% ±3.3 |
| 22 | Muse Spark Open | xhigh | 85.6% ±3.3 |