Refusal to assist with AI safety R&D (Anthropic)
Models receive single-turn coding requests for legitimate AI safety research with user context. Score is the refusal rate, where lower is better.
In-house Anthropic evaluation. The tasks 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.
Claude Fable 5 leads with 3.10% among models released by Jun 2026
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 Jun 2026, 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 15 more
- 1 Claude Fable 5 3.10%
- 2 Claude Opus 4.6 14%
- 3 Claude Mythos Preview 18.7%
- 4 Claude Opus 4.7 33.1%
- 5 Claude Sonnet 4.6 38%
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude Fable 5 | 3.10% | Jun 2026 | Model card | |
| 2 | Claude Opus 4.6 | 14% | Feb 2026 | Model card | |
| 3 | Claude Mythos Preview | 18.7% | Apr 2026 | Model card | |
| 4 | Claude Opus 4.7 | 33.1% | Apr 2026 | Model card | |
| 5 | Claude Sonnet 4.6 | 38% | Feb 2026 | Model card |
2 models have held number one since its release in Feb 2026
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 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.
6 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 |
|---|---|---|---|---|
| 9 Jun 2026 | Claude Fable 5 system card | 3.1 | Refusal to assist with AI safety R&D, safety-research refusal rate with extended thinking, 0.031 printed on chart, lower is better (Figure 6.3.4.A) | Alignment |
| 28 May 2026 | Claude Opus 4.8 system card | Refusal to assist with AI safety R&D, refusal rate with and without extended thinking, chart without printed values. Text says the lowest of recent models (6.3.4, Figure 6.3.4.A) | Alignment | |
| 16 Apr 2026 | Claude Opus 4.7 system card | 33.1 | Alignment-research refusals, refusal score 0.331 with thinking (16k budget), lower is better. 0.338 without thinking (6.3.4, Figure 6.3.4.A) | Alignment |
| 7 Apr 2026 | Claude Mythos Preview system card | 18.7 | Alignment Research Refusals Rate with thinking, 0.187 printed on chart (text says 18%), lower is better; 0.142 without thinking (4.3.4, Figure 4.3.4.A) | Alignment |
| 17 Feb 2026 | Claude Sonnet 4.6 system card | 38 | Alignment research refusals rate, thinking 0.380; no thinking 0.441 (Figure 4.6.1.A) | Alignment |
| 5 Feb 2026 | Claude Opus 4.6 system card | 14 | Refusal rate on reasonable AI safety research tasks, lower is better, stated in text (6.3.6, Figure 6.3.6.A) | Alignment |
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%