Benchmarks
347 sources · 1,258 models × 3,038 benchmarks · 158,912 scoresGoodhart Labs' HoneyBench measures specification gaming: each of its 20 agentic tasks in seven domains (medical, ML, administrative, coding, engineering, multimodal, and puzzles and games) pairs a hard objective with a faulty scoring setup that seems to let the agent score well while defeating the task's purpose. A run counts as a hack when its final submission exploits the flaw or the agent reaches proscribed information or resources, and the score is the share of scored runs hacked, pooled over the tasks a model was run on; lower is better. Diagnostic: provider-refused runs are left out of the rate and models run on different task sets.
No signal or difficulty estimate. This benchmark category is excluded from the capability model and its validation studies.
| # | Model | Setting | Score | Source |
|---|---|---|---|---|
| 1 | Claude Fable 5.2 | not reported | 11.9 | |
| 2 | Claude Opus 5.5 | not reported | 31.6 | |
| 3 | Claude Sonnet 5.5 | not reported | 41.6 | |
| 4 | GPT-6 Astra | not reported | 50.8 | |
| 5 | Inkling 2 Small | 0.99 | 52.6 | |
| 6 | GPT-6.1 Sol | not reported | 59.5 | |
| 7 | Claude Fable 5.1 | not reported | 61.8 | |
| 8 | Gemini 4 Argon | not reported | 63.6 | |
| 9 | Step 5 Preview | — | 64.9 | |
| 10 | Kimi K3.1 Pro | not reported | 71.8 | |
| 11 | DeepSeek V4.1 Flash 1017 | not reported | 76.4 | |
| 12 | GLM-5.3 | not reported | 76.8 | |
| 13 | Kimi K3 | not reported | 78.5 | |
| 14 | Gemini 3.8 Flash | not reported | 80.5 | |
| 15 | Muse Wonder | not reported | 81.0 | |
| 16 | Muse Spark 1.3 | not reported | 83.6 | |
| 17 | Grok 4.7 | not reported | 85.1 | |
| 18 | Muse Spark Open | not reported | 85.6 |
Benchmarks outside its family whose model rankings track it most reliably (Spearman ρ over at least 10 shared models).