Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
Loading page…
Preferred languages
More languages supported by Google Translate
Open the selector to load the current language catalog.
Translations are provided by Google Translate.
external benchmark
HoneyBench v0.2 (Goodhart Labs). Every model the catalog carries a published HoneyBench value for, ranked by that value.
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. 21 models in the ModelChoice catalog have a published value.
Each point is a model with a published HoneyBench value and complete token prices, on the 75% input / 25% output blend. Frontier models are spaced evenly along the cost axis.
21 models with a published HoneyBench value, ordered by that value, highest first. Models the source has not scored on this benchmark are not listed — they are unmeasured, not last.
| Rank | Model | Provider | Reward-hacking rate (%, lower is better) |
|---|---|---|---|
| 1 | Meta | 85.6 | |
| 2 | xAI | 85.1 | |
| 3 | Mistral | 83.6 |
What the benchmark measures and how its value is published, from the source’s own record.
HoneyBench (HoneyBench v0.2 (Goodhart Labs)) is an external benchmark published by HoneyBench v0.2. 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. Its evaluation set is described by the source as 20 private honeypot tasks across 7 domains with exploitable evaluators. The source rates its difficulty as agentic alignment and specification gaming.
Each model's result is published by HoneyBench v0.2 as reward-hacking rate (%, lower is better). ModelChoice prints that value to 1 decimal place, exactly as published, and never converts or rescales it. This page ranks the 21 models that carry a published value, highest value first; a model the source never scored is left out rather than ranked at zero. The source does not state whether a higher value is the better result, so neither does this page.
| 3 | Meta | 83.6 |
| 5 | Meta | 81.0 |
| 6 | 80.5 |
| 7 | Moonshot AI | 78.5 |
| 8 | Z.AI | 76.8 |
| 9 | DeepSeek | 76.4 |
| 10 | Moonshot AI | 71.8 |
| 11 | DeepSeek | 71.4 |
| 12 | Alibaba | 66.2 |
| 13 | StepFun | 64.9 |
| 14 | 63.6 |
| 15 | Anthropic | 61.8 |
| 16 | OpenAI | 59.5 |
| 17 | TMInkling 2 Small | Thinking Machines Lab | 52.6 |
| 18 | OpenAI | 50.8 |
| 19 | Anthropic | 41.6 |
| 20 | Anthropic | 31.6 |
| 21 | Anthropic | 11.9 |
Evidence key: Observed
Rows are ordered by the value HoneyBench v0.2 published, highest first. The source does not state whether a higher value is the better result, so this page does not either: for a benchmark that measures a rate of failure, read the table from the bottom.