RefusalBench (Nous Research)
Models get requests in scenarios that closed and open models commonly refuse. The score is the percent of questions the model answers, averaged over 5 trials.
Nous introduced it with Hermes 4 and keeps a RefusalBench leaderboard that it cites for Hermes 4.3. Higher means fewer refusals, which Nous treats as better steerability.
- Level
- Everyday Most adults, without training
- Scored
- Graded by a model A language model scores each response, usually against a rubric.
- Answers
- Open-ended Free-form text, documents, designs or media judged as a whole.
Hermes 4.3 36B leads with 74.6%
Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.
- 1 Hermes 4.3 36B 74.6%
- 2 Hermes 4 405B 57.1%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Hermes 4.3 36B | 74.6% | Dec 2025 | Model card | |
| 2 | Hermes 4 405B | 57.1% | Aug 2025 | Model card |
2 models have held number one since its release in Aug 2025
Lab claims count from the day the card came out, independent runs from the model’s release.
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Labs headline it less often now, last in Dec 2025
Launch posts from every lab that has listed it, one tick per post. Tall colored ticks include it, and short gray ones leave it out.
| Date | Launch post | Score | As the post writes it |
|---|---|---|---|
| 3 Dec 2025 | Hermes 4.3 36B | 74.6 | RefusalBench, % of questions answered, non-reasoning (reasoning 72.29), average of 5 trials |
| 26 Aug 2025 | Hermes 4 405B | 57.1 | RefusalBench (reasoning) |
One system card reports it
Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.
| Date | Document | Score | As the document writes it | Section |
|---|---|---|---|---|
| 25 Aug 2025 | Hermes 4 405B technical report | 57.1 | RefusalBench, average of 5 runs, higher means fewer refusals (also Figure 4), reasoning mode (Table 3). Non-reasoning 43.2 | Safety |
More in Safety and alignment
- FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
- StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is.
Inkling 98.6% - AgentDojo In use Agents do tool-use tasks while tools and data carry prompt injections. Attack success rate counts hijacked runs. Grok 4 Fast 0%
- MASK In use Measures whether a model contradicts its own elicited beliefs under pressure to lie. Honesty separated from accuracy.
Muse Spark Open 98.1% - Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. GPT-6 Astra 8.50%
- Agentic misalignment suite (Anthropic) In use Agents in simulated company settings face threats or goal conflicts that invite blackmail, leaking or sabotage. Score is the misaligned action rate. Claude Sonnet 4.5 0%