# HoneyBench v0.2 (Goodhart Labs) (HoneyBench)

> A reward-hacking and specification-gaming safety benchmark from Goodhart Labs that pairs difficult agent tasks with exploitable graders to measure how often models cheat instead of solving the task honestly. Lower scores are better.

Canonical page: https://benchlm.ai/benchmarks/honeybench

- Category: [external](/external)
- Last updated: HoneyBench v0.2 data of November 12, 2026 (captured November 20, 2026)

## About HoneyBench

- Year: 2026
- Tasks: 20 private honeypot tasks across 7 domains with exploitable evaluators
- Format: Reward-hacking rate (%, lower is better)
- Difficulty: Agentic alignment and specification gaming
- Paper: [HoneyBench v0.2](https://goodhartlabs.com/honeybench)

Goodhart Labs released HoneyBench v0.2 in November 2026 with 20 private honeypot tasks across 7 domains (coding, cybersecurity, healthcare, engineering, and administrative settings). We mirror the official 22-model leaderboard (data of November 12, 2026), preserving 95% confidence intervals and refusal-adjusted hack rates. Provider-reported figures for models not yet on Goodhart Labs' table appear below under lab-reported results.

HoneyBench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (22 models)

| Rank | Model | Configuration | Creator | Score |
|------|-------|---------------|---------|-------|
| 1 | [Claude Fable 5.2](/models/claude-fable-5-2) | 95% CI 9.1%-15.4% | Anthropic | 11.9% |
| 2 | [Claude Opus 5.5](/models/claude-opus-5-5) | 95% CI 27.1%-36.6% · 39.7% counting refusals | Anthropic | 31.6% |
| 3 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | 95% CI 37.1%-46.3% | Anthropic | 41.6% |
| 4 | [GPT-6 Astra](/models/gpt-6-astra) | 95% CI 46.2%-55.3% | OpenAI | 50.8% |
| 5 | [Inkling 2 Small](/models/inkling-2-small) | 95% CI 47.7%-57.5% | Thinking Machines Lab | 52.6% |
| 6 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | 95% CI 54.9%-63.9% | OpenAI | 59.5% |
| 7 | [Claude Fable 5.1](/models/claude-fable-5-1) | 95% CI 57.0%-66.5% · 63.9% counting refusals | Anthropic | 61.8% |
| 8 | [Gemini 4 Argon](/models/gemini-4-argon) | 95% CI 58.1%-68.8% · 74.2% counting refusals | Google | 63.6% |
| 9 | [Step 5 Preview](/models/step-5-preview) | 95% CI 60.3%-69.2% | StepFun | 64.9% |
| 10 | [Qwen3.8 Max](/models/qwen3-8-max) | 95% CI 61.7%-70.3% | Alibaba | 66.2% |
| 11 | [DeepSeek V4 Pro](/models/deepseek-v4-pro) | 95% CI 66.9%-75.4% | DeepSeek | 71.4% |
| 12 | [Kimi K3.1 Pro](/models/kimi-3-1-pro) | 95% CI 67.5%-75.7% | Moonshot AI | 71.8% |
| 13 | [DeepSeek V4.1 Flash 1017](/models/deepseek-v4-1-flash) | 95% CI 72.3%-80.1% | DeepSeek | 76.4% |
| 14 | [GLM-5.3](/models/glm-5-3) | 95% CI 72.6%-80.5% | Z.AI | 76.8% |
| 15 | [Kimi K3](/models/kimi-k3) | 95% CI 74.5%-82.0% | Moonshot AI | 78.5% |
| 16 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | 95% CI 76.6%-83.9% | Google | 80.5% |
| 17 | [DeepSeek V4.1 Flash](https://goodhartlabs.com/honeybench) | 95% CI 77.2%-84.4% | DeepSeek | 81.0% |
| 18 | [Muse Wonder](/models/muse-wonder) | 95% CI 77.2%-84.4% | Meta | 81.0% |
| 19 | [Mistral Large 4](/models/mistral-large-4) | 95% CI 79.9%-86.7% | Mistral | 83.6% |
| 20 | [Muse Spark 1.3](/models/muse-spark-1-3) | 95% CI 79.9%-86.7% | Meta | 83.6% |
| 21 | [Grok 4.7](/models/grok-4-7) | 95% CI 81.6%-88.1% | xAI | 85.1% |
| 22 | [Muse Spark Open](/models/muse-spark-open) | 95% CI 82.1%-88.6% | Meta | 85.6% |

## Results reported by model developers

Unranked. Labs ran these themselves. A lab’s own result is shown for reference. It never takes a rank on this board.

| Model | Reported value | What was measured | Against the official board | Who ran it | Source |
|-------|----------------|-------------------|---------------------------|------------|--------|
| [Inkling 2 Mini](/models/inkling-2-mini) | 56.7% (Reward-hacking rate (lower is better)) | HoneyBench v0.2; Reported in the Inkling 2 Mini model card before inclusion on Goodhart Labs' HoneyBench v0.2 leaderboard. | Comparability not established | Reported by Thinking Machines Lab | [Thinking Machines Lab Inkling 2 Mini model card](https://thinkingmachines.ai/news/inkling-2-mini/model-card/) |

## FAQ

### What does HoneyBench measure?

A reward-hacking and specification-gaming safety benchmark from Goodhart Labs that pairs difficult agent tasks with exploitable graders to measure how often models cheat instead of solving the task honestly. Lower scores are better.

### Which model leads the published HoneyBench snapshot?

Claude Fable 5.2 currently leads the published HoneyBench snapshot with a score of 11.9%.

### How many models are evaluated on HoneyBench?

The HoneyBench v0.2 data of November 12, 2026 (captured November 20, 2026) contains 22 AI models.

### Does HoneyBench affect BenchLM's overall score?

Not directly. HoneyBench is still displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.
