Benchmark Almanac

LiveSecBench

Claude Haiku 4.5 91.4 leads since November 2025
Not in model cards

Models answer Chinese-language prompts that probe safety across areas such as legality, ethics, factuality and privacy. The paper's first edition turned a judge's pairwise comparisons into Elo ratings. The leaderboard's March 2026 edition gives each model an overall score from 0 to 100. The question bank is updated in dated editions.

The questions are not published, only the harness and results. The paper's first edition (v251030) scored 18 models. The leaderboard now shows a March 2026 edition. 29 GitHub stars and 4 HF paper upvotes, so traction rests on the maintained leaderboard.

Released
November 2025
Kind
Refusing harmful requests
Level
Everyday
Scored
Graded by a model
Answers
Open-ended
Tasks
Kept private
Strongest model tested
Gemini 3.1 Pro #39 on the capability index
Last model added
February 2026
Scores refreshed
November 12th, 2026

In the authors’ words

We introduce LiveSecBench, a continuously updated safety benchmark specifically for Chinese-language LLM application scenarios. LiveSecBench constructs a high-quality and unique dataset through a pipeline that combines automated generation with human verification. By periodically releasing new versions to expand the dataset and update evaluation metrics, LiveSecBench provides a robust and up-to-date standard for AI safety. In this report, we introduce our second release v251215, which evaluates across five dimensions (Public Safety, Fairness & Bias, Privacy, Truthfulness, and Mental Health Safety.) We evaluate 57 representative LLMs using an ELO rating system, offering a leaderboard of the current state of Chinese LLM safety. The result is available at https://livesecbench.intokentech.cn/.

Abstract of LiveSecBench: A Dynamic and Event-Driven Safety Benchmark for Chinese Language Model Applications (arXiv:2511.02366, CC0)
43
Entries on its own leaderboard, 34 models we track

Claude Haiku 4.5 leads with 91.4 among models released by February 2026

Best score per entry from the benchmark's own leaderboard, read on November 12th, 2026. Where a model ran with several settings or harnesses, we keep its best. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders.

The newest model with a score here came out in February 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

43 entries. Scroll sideways for the rest.

  1. 91.4 Claude Haiku 4.5
  2. 86.0 Claude Sonnet 4.6
  3. 84.7 GPT-5.2
  4. 84.3 Qwen3.5-Plus
  5. 81.5 Qwen3.5 397B-A17B
  6. 79.2 spark-x
  7. 74.8 Kimi K2.5
  8. 70.8 doubao-seed-1-6-250615
  9. 69.2 Qwen3-235B-A22B-Instruct
  10. 66.7 MiniMax-M2
  11. 66.6 gpt-oss-120b
  12. 63.6 Intern-S1-Pro
  13. 63.0 Doubao-Seed-2.0 Pro
  14. 61.6 MiniMax-M2.5
  15. 58.2 Gemini 3.1 Pro
  16. 57.2 MiMo-V2-Flash
  17. 57.1 LongCat-Flash-Chat
  18. 56.7 GLM-5
  19. 56.2 DeepSeek-V3.2
  20. 55.2 DeepSeek-R1
  21. 52.1 Step 3.5 Flash
  22. 44.9 GLM-4.6
  23. 42.4 gemini-2.5-flash (snapshot unspecified)
  24. 42.1 SenseChat-Turbo-1202
  25. 40.5 Ling-2.5-1T
  26. 40.0 Ernie-5.0-Preview-1022 (snapshot not cataloged)
  27. 38.9 Step 3
  28. 38.9 Llama 3.3 70B Instruct
  29. 35.6 Kimi K2
  30. 32.7 Grok 4.1 Fast
  31. 32.6 hunyuan-t1-20250822 (snapshot not cataloged)
  32. 31.5 Intern-S1
  33. 30.3 seed-oss-36b-instruct
  34. 29.7 Mistral Large 2
  35. 28.4 Mistral Large 3
  36. 28.4 Grok-3 mini
  37. 28.2 Llama 4 Maverick
  38. 24.0 SenseNova-V6-5-Turbo
  39. 23.7 ERNIE-4.5-21B-A3B-Thinking
  40. 23.0 GPT-4.1 mini
  41. 18.1 DeepSeek-V3
  42. 16.9 Hunyuan-A13B-Instruct
  43. 14.3 ascend-tribe/pangu-pro-moe
  1. 1 Claude Haiku 4.5 91.4
  2. 2 Claude Sonnet 4.6 86.0
  3. 3 GPT-5.2 84.7
  4. 4 Qwen3.5-Plus 84.3
  5. 5 Qwen3.5 397B-A17B 81.5
  6. 6 spark-x 79.2
  7. 7 Kimi K2.5 74.8
  8. 8 doubao-seed-1-6-250615 70.8
  9. 9 Qwen3-235B-A22B-Instruct Jul 2025 69.2
  10. 10 MiniMax-M2 66.7
  11. 11 gpt-oss-120b 66.6
  12. 12 Intern-S1-Pro 63.6
  13. 13 Doubao-Seed-2.0 Pro 63.0
  14. 14 MiniMax-M2.5 61.6
  15. 15 Gemini 3.1 Pro 58.2
  16. 16 MiMo-V2-Flash 57.2
  17. 17 LongCat-Flash-Chat 57.1
  18. 18 GLM-5 56.7
  19. 19 DeepSeek-V3.2 56.2
  20. 20 DeepSeek-R1 May 2025 55.2
  21. 21 Step 3.5 Flash 52.1
  22. 22 GLM-4.6 44.9
  23. 23 gemini-2.5-flash (snapshot unspecified) 42.4
  24. 24 SenseChat-Turbo-1202 42.1
  25. 25 Ling-2.5-1T 40.5
  26. 26 Ernie-5.0-Preview-1022 (snapshot not cataloged) 40.0
  27. 27 Step 3 38.9
  28. 28 Llama 3.3 70B Instruct 38.9
  29. 29 Kimi K2 Jul 2025 35.6
  30. 30 Grok 4.1 Fast 32.7
  31. 31 hunyuan-t1-20250822 (snapshot not cataloged) 32.6
  32. 32 Intern-S1 31.5
  33. 33 seed-oss-36b-instruct 30.3
  34. 34 Mistral Large 2 Nov 2024 29.7
  35. 35 Mistral Large 3 28.4
  36. 36 Grok-3 mini 28.4
  37. 37 Llama 4 Maverick 28.2
  38. 38 SenseNova-V6-5-Turbo 24.0
  39. 39 ERNIE-4.5-21B-A3B-Thinking 23.7
  40. 40 GPT-4.1 mini 23.0
  41. 41 DeepSeek-V3 Mar 2025 18.1
  42. 42 Hunyuan-A13B-Instruct 16.9
  43. 43 ascend-tribe/pangu-pro-moe 14.3

Source: Official LiveSecBench leaderboard

#EntryScoreSettingDateSource
1 Claude Haiku 4.5 91.4 Oct 2025 Official leaderboard
2 Claude Sonnet 4.6 86.0 Feb 2026 Official leaderboard
3 GPT-5.2 84.7 Dec 2025 Official leaderboard
4 Qwen3.5-Plus 84.3 Feb 2026 Official leaderboard
5 Qwen3.5 397B-A17B 81.5 Feb 2026 Official leaderboard
6 spark-x 79.2 Mar 2026 Official leaderboard
7 Kimi K2.5 74.8 Jan 2026 Official leaderboard
8 doubao-seed-1-6-250615 70.8 Mar 2026 Official leaderboard
9 Qwen3-235B-A22B-Instruct Jul 2025 69.2 Jul 2025 Official leaderboard
10 MiniMax-M2 66.7 Oct 2025 Official leaderboard
11 gpt-oss-120b 66.6 Aug 2025 Official leaderboard
12 Intern-S1-Pro 63.6 Feb 2026 Official leaderboard
13 Doubao-Seed-2.0 Pro 63.0 Feb 2026 Official leaderboard
14 MiniMax-M2.5 61.6 Feb 2026 Official leaderboard
15 Gemini 3.1 Pro 58.2 Feb 2026 Official leaderboard
16 MiMo-V2-Flash 57.2 Dec 2025 Official leaderboard
17 LongCat-Flash-Chat 57.1 Aug 2025 Official leaderboard
18 GLM-5 56.7 Feb 2026 Official leaderboard
19 DeepSeek-V3.2 56.2 Dec 2025 Official leaderboard
20 DeepSeek-R1 May 2025 55.2 May 2025 Official leaderboard
21 Step 3.5 Flash 52.1 Feb 2026 Official leaderboard
22 GLM-4.6 44.9 Sep 2025 Official leaderboard
23 gemini-2.5-flash (snapshot unspecified) 42.4 Mar 2026 Official leaderboard
24 SenseChat-Turbo-1202 42.1 Mar 2026 Official leaderboard
25 Ling-2.5-1T 40.5 Mar 2026 Official leaderboard
26 Ernie-5.0-Preview-1022 (snapshot not cataloged) 40.0 Mar 2026 Official leaderboard
27 Step 3 38.9 Jul 2025 Official leaderboard
28 Llama 3.3 70B Instruct 38.9 Dec 2024 Official leaderboard
29 Kimi K2 Jul 2025 35.6 Jul 2025 Official leaderboard
30 Grok 4.1 Fast 32.7 Nov 2025 Official leaderboard
31 hunyuan-t1-20250822 (snapshot not cataloged) 32.6 Mar 2026 Official leaderboard
32 Intern-S1 31.5 Jul 2025 Official leaderboard
33 seed-oss-36b-instruct 30.3 Aug 2025 Official leaderboard
34 Mistral Large 2 Nov 2024 29.7 Nov 2024 Official leaderboard
35 Mistral Large 3 28.4 Dec 2025 Official leaderboard
36 Grok-3 mini 28.4 Apr 2025 Official leaderboard
37 Llama 4 Maverick 28.2 Apr 2025 Official leaderboard
38 SenseNova-V6-5-Turbo 24.0 Mar 2026 Official leaderboard
39 ERNIE-4.5-21B-A3B-Thinking 23.7 Sep 2025 Official leaderboard
40 GPT-4.1 mini 23.0 Apr 2025 Official leaderboard
41 DeepSeek-V3 Mar 2025 18.1 Mar 2025 Official leaderboard
42 Hunyuan-A13B-Instruct 16.9 Jun 2025 Official leaderboard
43 ascend-tribe/pangu-pro-moe 14.3 Mar 2026 Official leaderboard

Number one over time

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

Anthropic
Own leaderboard Not tested since Feb 2026
Claude Haiku 4.5 91.4

Source: Official LiveSecBench leaderboard

Scores over time

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

On the benchmark’s own leaderboard
0.00 20.0 40.0 60.0 80.0 100.0 2025 2026 0.00 20.0 40.0 60.0 80.0 100.0 2025 2026

Source: Official LiveSecBench leaderboard

No lab model card reports it yet

We read the headline results of 522 launch posts from 52 labs, and none of them list this benchmark.

More in Safety and alignment

All 465 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%