Challenging Refusal Evaluation
Models answer a harder set of disallowed content prompts across policy categories. An autograder scores the not_unsafe rate.
In-house OpenAI evaluation. The prompts are not public.
- Level
- Expert Specialists with years of training, often over hours of work
- 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.
o1-preview leads with 93.4% among models released by Apr 2025
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.
The newest model with a score here came out in Apr 2025, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6 Astra
- Claude Opus 5.5
- GPT-6.1 Terra
- GPT-6.1 Sol
- and 16 more
- 1 o1-preview 93.4%
- 2 o1-mini 93.2%
- 3 o1 92%
- 4 OpenAI Computer-Using Agent (CUA) 92%
- 5 o3 92%
- 6 o3-mini 90%
- 7 o4-mini 90%
- 8 OpenAI deep research 89%
- 9 GPT-4.5 85%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | o1-preview | 93.4% | Sep 2024 | Model card | |
| 2 | o1-mini | 93.2% | Sep 2024 | Model card | |
| 3 | o1 | 92% | Dec 2024 | Model card | |
| 4 | OpenAI Computer-Using Agent (CUA) | 92% | Jan 2025 | Model card | |
| 5 | o3 | 92% | Apr 2025 | Model card | |
| 6 | o3-mini | 90% | Jan 2025 | Model card | |
| 7 | o4-mini | 90% | Apr 2025 | Model card | |
| 8 | OpenAI deep research | 89% | Feb 2025 | Model card | |
| 9 | GPT-4.5 | 85% | Feb 2025 | Model card |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
The best score passed 85% in Sep 2024
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Labs report it only in system cards
We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark. Some system cards do, below.
9 system cards report it, from 1 lab
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 |
|---|---|---|---|---|
| 16 Apr 2025 | o4-mini system card | 90 | Challenging refusal evaluation, aggregate not_unsafe (Table 2) | Safety |
| 16 Apr 2025 | o3 system card | 92 | Challenging refusal evaluation, aggregate not_unsafe (Table 2) | Safety |
| 27 Feb 2025 | GPT-4.5 system card | 85 | Challenging Refusal Evaluation, not_unsafe (Table 1) | Safety |
| 25 Feb 2025 | OpenAI deep research system card | 89 | Disallowed content evaluations (challenging), not_unsafe_challenging 0.89 (Table 4) | Safety |
| 31 Jan 2025 | o3-mini system card | 90 | Challenging Refusal Evaluation, not_unsafe (Table 1) | Safety |
| 23 Jan 2025 | OpenAI Computer-Using Agent (CUA) system card | 92 | Challenging refusal evaluation, not_unsafe (Table 3, categories in Table 5) | Safety |
| 5 Dec 2024 | o1 system card | 92 | Challenging Refusal Evaluation, not_unsafe (Table 1) | Safety |
| 12 Sep 2024 | o1-preview system card | 93.4 | Challenging Refusal Evaluation, not_unsafe (Table 1) | Safety |
| 12 Sep 2024 | o1-mini system card | 93.2 | Challenging Refusal Evaluation, not_unsafe (Table 1) | 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%