Benchmark Almanac

Single-turn harmful request evaluations (Anthropic)

Claude Opus 4.5 99.8% leads since Nov 2025 Only in system cards

Models answer single-turn requests that violate the usage policy. Score is the harmless response rate.

In-house Anthropic evaluation. The tasks are not public.

Made by Anthropic Released May 2025 Refusing harmful requests Tasks kept private Last in a system card Oct 2026 Last model added Oct 2026
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.
18
Models scored in lab cards
18
System cards report it, from 1 lab

Claude Opus 4.5 leads with 99.8%

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.

18 models. Scroll sideways for the rest.

  1. 99.8 Claude Opus 4.5
  2. 99.6 Claude Opus 4.6
  3. 99.4 Claude Haiku 4.5
  4. 99.4 Claude Sonnet 4.6
  5. 99.3 Claude Sonnet 4.5
  6. 99.0 Claude Sonnet 4
  7. 98.8 Claude Opus 4.1
  8. 98.4 Claude Opus 4
  9. 98.4 Claude Haiku 5.5
  10. 98.0 Claude Opus 4.7
  11. 98.0 Claude Opus 4.8
  12. 97.8 Claude Mythos Preview
  13. 96.9 Claude Fable 5
  14. 96.7 Claude Sonnet 5
  15. 96.3 Claude Opus 5
  16. 95.6 Claude Sonnet 5.5
  17. 94.7 Claude Fable 5.1
  18. 94.5 Claude Opus 5.5
  1. 1 Claude Opus 4.5 99.8%
  2. 2 Claude Opus 4.6 99.6%
  3. 3 Claude Haiku 4.5 99.4%
  4. 4 Claude Sonnet 4.6 99.4%
  5. 5 Claude Sonnet 4.5 99.3%
  6. 6 Claude Sonnet 4 99%
  7. 7 Claude Opus 4.1 98.8%
  8. 8 Claude Opus 4 98.4%
  9. 9 Claude Haiku 5.5 98.4%
  10. 10 Claude Opus 4.7 98%
  11. 11 Claude Opus 4.8 98%
  12. 12 Claude Mythos Preview 97.8%
  13. 13 Claude Fable 5 96.9%
  14. 14 Claude Sonnet 5 96.7%
  15. 15 Claude Opus 5 96.3%
  16. 16 Claude Sonnet 5.5 95.6%
  17. 17 Claude Fable 5.1 94.7%
  18. 18 Claude Opus 5.5 94.5%
#ModelScoreSettingDateSource
1 Claude Opus 4.5 99.8% Nov 2025 Model card
2 Claude Opus 4.6 99.6% Feb 2026 Model card
3 Claude Haiku 4.5 99.4% Oct 2025 Model card
4 Claude Sonnet 4.6 99.4% Feb 2026 Model card
5 Claude Sonnet 4.5 99.3% Sep 2025 Model card
6 Claude Sonnet 4 99% May 2025 Model card
7 Claude Opus 4.1 98.8% Aug 2025 Model card
8 Claude Opus 4 98.4% May 2025 Model card
9 Claude Haiku 5.5 98.4% Oct 2026 Model card
10 Claude Opus 4.7 98% Apr 2026 Model card
11 Claude Opus 4.8 98% May 2026 Model card
12 Claude Mythos Preview 97.8% Apr 2026 Model card
13 Claude Fable 5 96.9% Jun 2026 Model card
14 Claude Sonnet 5 96.7% Jun 2026 Model card
15 Claude Opus 5 96.3% Jul 2026 Model card
16 Claude Sonnet 5.5 95.6% Sep 2026 Model card
17 Claude Fable 5.1 94.7% Sep 2026 Model card
18 Claude Opus 5.5 94.5% Sep 2026 Model card

4 models have held number one since its release in May 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Lab claims Passed 85% in May 2025, by Claude Sonnet 4
Claude Opus 4.5 99.8%

The best score passed 85% in May 2025

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Claude Opus 4.5 Claude Sonnet 4 0% 20% 40% 60% 80% 100% 2025 2026 Near the ceiling (85%) Claude Opus 4.5

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.

18 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.

DateDocumentScoreAs the document writes itSection
7 Oct 2026 Claude Haiku 5.5 system card 98.4 Single-turn harmful request evaluations, overall harmless response rate, API without system prompt (claude.ai 99.71%) (Table 4.1.1.A) Safety
28 Sep 2026 Claude Sonnet 5.5 system card 95.6 Single-turn harmful requests, overall harmless response rate, API without system prompt (claude.ai 99.40%) (Table 4.1.1.A) Safety
22 Sep 2026 Claude Opus 5.5 system card 94.5 Single-turn harmful requests, overall harmless response rate, API without a system prompt, 16 policy areas, 7 languages (claude.ai 99.51%) (Table 4.1.1.A) Safety
1 Sep 2026 Claude Fable 5.1 system card 94.7 Single-turn harmful request evaluations, overall harmless response rate, API without a system prompt, 7 languages; claude.ai with system prompt 99.53% (Table 4.1.1.A) Safety
24 Jul 2026 Claude Opus 5 system card 96.3 Single-turn harmful requests, overall harmless response rate, API without a system prompt, 16 policy areas in 7 languages (claude.ai 98.54%) (Table 4.1.1.A) Safety
30 Jun 2026 Claude Sonnet 5 system card 96.7 Single-turn harmful requests, overall harmless response rate, API without a system prompt, 16 policy areas, 7 languages (claude.ai 99.20%) (Table 4.1.1.A) Safety
9 Jun 2026 Claude Fable 5 system card 96.9 Single-turn harmful request evaluations, overall harmless response rate, API without a system prompt, 16 policy areas in 7 languages; claude.ai 98.51% (Table 4.1.1.A) Safety
28 May 2026 Claude Opus 4.8 system card 98 Single-turn harmful requests, overall harmless response rate, API without a system prompt, 16 policy areas, 7 languages (claude.ai 99.17%) (Table 4.1.1.A) Safety
16 Apr 2026 Claude Opus 4.7 system card 98 Single-turn violative request evaluations, overall harmless response rate, 7 languages. Without thinking 98.84, with adaptive thinking 97.12 (Table 4.1.1.A) Safety
7 Apr 2026 Claude Mythos Preview system card 97.8 Single-turn violative request evaluations, overall harmless response rate, 7 languages (Table 8.1.1.1.A); 98.33 without thinking, 97.35 with adaptive thinking Safety
17 Feb 2026 Claude Sonnet 4.6 system card 99.4 Single-turn violative request evaluations, overall harmless response rate, 7 languages (Table 3.1.1.A) Safety
5 Feb 2026 Claude Opus 4.6 system card 99.6 Single-turn violative request evaluations, overall harmless response rate, 7 languages (Table 3.1.1.A); default 99.53, extended thinking 99.74 Safety
24 Nov 2025 Claude Opus 4.5 system card 99.8 Single-turn violative request evaluations, overall harmless response rate, 6 languages (Table 3.1.1.A; default 99.70%, extended thinking 99.85%) Safety
15 Oct 2025 Claude Haiku 4.5 system card 99.4 Single-turn violative request evaluations, overall harmless response rate (Table 2.1.1.A) Safety
29 Sep 2025 Claude Sonnet 4.5 system card 99.3 Single-turn violative request evaluations, overall harmless response rate 99.29% (± 0.22%), English only. Standard thinking 99.16%, extended thinking 99.43% (Table 2.1.1.A) Safety
5 Aug 2025 Claude Opus 4.1 system card addendum 98.8 Single-turn violative request evaluations, English only, overall harmless response rate 98.76% (± 0.29%). Standard thinking 98.45%, extended thinking 99.06% (Table 2.1.A) Safety
22 May 2025 Claude Sonnet 4 system card 99 Single-turn violative request evaluations, overall harmless response rate 98.99% (± 0.23%). Standard thinking 98.59%, extended thinking 99.40% (Table 2.1.A) Safety
22 May 2025 Claude Opus 4 system card 98.4 Single-turn violative request evaluations, overall harmless response rate 98.43% (± 0.30%), model without ASL-3 safeguards. Standard thinking 97.92%, extended thinking 98.94%, with ASL-3 safeguards 98.76% (Table 2.1.A) Safety

More in Safety and alignment

All 444 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%