Harmlessness human feedback win rate (Anthropic)
Raters probe models in adversarial chats and pick the more harmless reply in head-to-head comparisons. Score is the win rate against a baseline model.
In-house Anthropic evaluation. A 50% win rate means equal to the baseline model, which changes between cards.
- Made by
- Anthropic
- Released
- June 2024
- Kind
- Refusing harmful requests
- Level
- Everyday
- Scored
- Ranked by votes
- Answers
- Open-ended
- Tasks
- Kept private
- Strongest model tested
- Claude 3.5 Sonnet (October 2024) #149 on the capability index
- Last in a system card
- October 2024
- Last model added
- October 2024
4 models have ratings here, all from their labs’ own posts
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. A rating ranks a model only against others rated in the same arena at the same time, so these don’t show who leads.
The newest model with a score here came out in October 2024, and nobody has reported a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- Claude Fable 5.2
-
Gemini 4 Carbon - GPT-6 Astra
- Claude Opus 5.5
- and 16 more
- 1 Claude 3 Sonnet 87%
- 2 Claude 3.5 Sonnet 52%
- 3 Claude 3.5 Sonnet (October 2024) 48%
- 4 Claude 3.5 Haiku 43%
Source: Labs’ model cards
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Claude 3 Sonnet | 87% | Mar 2024 | Lab model card | |
| 2 | Claude 3.5 Sonnet | 52% | Jun 2024 | Lab model card | |
| 3 | Claude 3.5 Sonnet (October 2024) | 48% | Oct 2024 | Lab model card | |
| 4 | Claude 3.5 Haiku | 43% | Oct 2024 | Lab model card |
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Labs’ model cards
Anthropic’s own number for Claude 3 Sonnet passed 85% in March 2024
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Labs’ model cards
Labs report it only in system cards
We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark. Some system cards do, below.
4 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 |
|---|---|---|---|---|
| Oct 22, 2024 | Claude 3.5 Sonnet (October 2024) model card | 48 | Human feedback win rate vs Claude 3.5 Sonnet baseline, Harmlessness (adversarial scenario) 48 (Figure 2) | Safety |
| Oct 22, 2024 | Claude 3.5 Haiku model card | 43 | Human feedback win rate vs Claude 3.5 Sonnet baseline, Harmlessness (adversarial scenario) 43 (Figure 2) | Safety |
| Jun 20, 2024 | Claude 3.5 Sonnet model card | 52 | Harmlessness, human preference win rate vs Claude 3 Opus baseline, adversarial scenarios (Figure 3) | Safety |
| Mar 4, 2024 | Claude 3 Sonnet model card | 87 | Human preference win rate vs Helpful-only baseline, Harmlessness (adversarial scenario) 87 (Figure 8) | 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% - 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. Gemini 4 Carbon 0%
- DistressBench In use Models reply as a chatbot to 718 simulated users in psychological crisis, in single-turn and multi-turn chats. Three judge models score replies against clinician-written rubrics.
Muse Spark 1.2 88.4% - Lakera B3 (Backbone Breaker Benchmark) In use Crowdsourced prompt-injection attacks hit backbone LLMs inside agent threat snapshots at three defense levels. The score is attack resistance, 100 minus the board's vulnerability percentage. Claude Opus 4.5 89%
- IHEval In use Models face conflicts between system, user and tool instructions. Reports a composite score for following the higher-priority instruction. GPT-4o 70%