Benchmark Almanac

CASE-Bench

Claude 3.5 Sonnet 90.1% leads since January 2025
Not in model cards

Models decide whether a request is safe in a given context. The same request appears in two contexts with different human judgments.

Made by
Guangzhi Sun and colleagues
Released
January 2025
Kind
Refusing harmful requests
Level
Everyday
Scored
Checked automatically
Answers
Multiple choice
Tasks
All public
Measures
Capability
Risk
Privacy
Language
English
In and out
Text → Text
Interaction
Single turn
Refresh
One fixed set
Run by
Anyone
In views
Content safety
Strongest model tested
Claude 3.5 Sonnet #158 on the capability index
Last model added
August 2024
Scores refreshed
November 20th, 2026
14
Entries in its paper, 11 models we track

Claude 3.5 Sonnet leads with 90.1% among models released by August 2024 Claude 3.5 Sonnet leads with 90.1% among models released by August 2024

Best score per entry from the benchmark's paper, read on November 20th, 2026. A paper’s table doesn’t add newer models. Where a model ran with several settings or harnesses, we keep its best. Every run is in the next chart. Names in gray aren't in our model list: fine-tunes, agent systems, products and older models. They don't count toward leaders. This view shows only models we track, each at its best.

The newest model with a score here came out in August 2024, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:

  1. 1 Claude 3.5 Sonnet Score 90.1%
  2. 2 DeepSeek-R1 (snapshot unspecified) Score 88.6%
  3. 3 Llama 3-70B Prob. 87.2%
  4. 4 Llama 3.1-8B Score 87%
  5. 5 Qwen2-72B Binary 86.4%
  6. 6 Mixtral 8x7B Score 83.1%
  7. 7 Mistral-7B-Instruct-v0.2 Prob. 82.7%
  8. 8 Dolphin-2.9-Llama3-70B Binary 82.6%
  9. 9 Llama 3-8B Prob. 82.4%
  10. 10 GPT-4o mini Binary 82%
  11. 11 Zephyr 7B Beta Prob. 81.6%
  12. 12 GPT-4o Aug 2024, Score 78.6%
  13. 13 Starling-LM-7B-alpha Prob. 73.4%
  14. 14 Llama-Guard-3-8B Context 60.1%

Source: The CASE-Bench paper

#EntryScoreSettingDateSource
1 Claude 3.5 Sonnet 90.1% Score Jun 2024 Official leaderboard
2 DeepSeek-R1 (snapshot unspecified) 88.6% Score Official leaderboard
3 Llama 3-70B 87.2% Prob. Apr 2024 Official leaderboard
4 Llama 3.1-8B 87% Score Jul 2024 Official leaderboard
5 Qwen2-72B 86.4% Binary Jun 2024 Official leaderboard
6 Mixtral 8x7B 83.1% Score Dec 2023 Official leaderboard
7 Mistral-7B-Instruct-v0.2 82.7% Prob. Dec 2023 Official leaderboard
8 Dolphin-2.9-Llama3-70B 82.6% Binary Official leaderboard
9 Llama 3-8B 82.4% Prob. Apr 2024 Official leaderboard
10 GPT-4o mini 82% Binary Jul 2024 Official leaderboard
11 Zephyr 7B Beta 81.6% Prob. Oct 2023 Official leaderboard
12 GPT-4o Aug 2024 78.6% Score Aug 2024 Official leaderboard
13 Starling-LM-7B-alpha 73.4% Prob. Nov 2023 Official leaderboard
14 Llama-Guard-3-8B 60.1% Context Official leaderboard
#ModelScoreSettingDateSource
1 Claude 3.5 Sonnet 90.1% Score Jun 2024 Official leaderboard
2 Llama 3-70B 87.2% Prob. Apr 2024 Official leaderboard
3 Llama 3.1-8B 87% Score Jul 2024 Official leaderboard
4 Qwen2-72B 86.4% Binary Jun 2024 Official leaderboard
5 Mixtral 8x7B 83.1% Score Dec 2023 Official leaderboard
6 Mistral-7B-Instruct-v0.2 82.7% Prob. Dec 2023 Official leaderboard
7 Llama 3-8B 82.4% Prob. Apr 2024 Official leaderboard
8 GPT-4o mini 82% Binary Jul 2024 Official leaderboard
9 Zephyr 7B Beta 81.6% Prob. Oct 2023 Official leaderboard
10 GPT-4o Aug 2024 78.6% Score Aug 2024 Official leaderboard
11 Starling-LM-7B-alpha 73.4% Prob. Nov 2023 Official leaderboard

Rankings by part

We read 2 parts from the benchmark’s paper: Safe recall and Unsafe recall. Each ranks models on one part of the test. Llama 3.1-8B leads Safe recall and Starling-LM-7B-alpha leads Unsafe recall. Arrows show how far a model moved from its overall rank among the models we track.

Safe recall

11 models we track have a score here.

  1. 1 Llama 3.1-8B ↑2 92.2%
  2. 2 Claude 3.5 Sonnet ↓1 90.7%
  3. 3 Llama 3-70B ↓1 89.2%
  4. 4 Qwen2-72B 76.9%
  5. 5 Zephyr 7B Beta ↑4 73.4%
  6. 6 Mistral-7B-Instruct-v0.2 72.7%
  7. 7 Mixtral 8x7B ↓2 71.7%
  8. 8 Llama 3-8B ↓1 68.9%
  9. 9 GPT-4o mini ↓1 67.4%
  10. 10 GPT-4o Aug 2024 58.1%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Claude 3.5 Sonnet
Binary
90.1%
Llama 3-70B
Score
87.2%
Llama 3.1-8B
Binary Prob.
87%
Qwen2-72B
Prob. Binary
86.4%
Mixtral 8x7B
Binary
83.1%
Mistral-7B-Instruct-v0.2
Binary Score Prob.
82.7%
Llama 3-8B
Score Prob.
82.4%
GPT-4o mini
Score
82%
Zephyr 7B Beta
Binary Prob.
81.6%
GPT-4o
Binary
78.6%
Starling-LM-7B-alpha
Binary Prob.
73.4%

Source: The CASE-Bench paper

ModelEffortHarnessScoreCost per taskOutput tokens
Claude 3.5 Sonnet Score 90.1%
Claude 3.5 Sonnet Binary 88.7%
Llama 3-70B Prob. 87.2%
Llama 3.1-8B Score 87%
Llama 3-70B Binary 86.6%
Qwen2-72B Binary 86.4%
Llama 3.1-8B Prob. 84.9%
Qwen2-72B Score 84.7%
Llama 3-70B Score 84.4%
Mixtral 8x7B Score 83.1%
Mixtral 8x7B Prob. 83.1%
Mistral-7B-Instruct-v0.2 Prob. 82.7%
Llama 3-8B Prob. 82.4%
GPT-4o mini Binary 82%
Mixtral 8x7B Binary 81.9%
Zephyr 7B Beta Prob. 81.6%
Qwen2-72B Prob. 81.6%
GPT-4o mini Score 80%
GPT-4o Aug 2024 Score 78.6%
GPT-4o Aug 2024 Binary 77%
Starling-LM-7B-alpha Prob. 73.4%
Llama 3-8B Binary 70.4%
Llama 3.1-8B Binary 70.3%
Llama 3-8B Score 70%
Mistral-7B-Instruct-v0.2 Score 66.3%
Zephyr 7B Beta Score 56.8%
Starling-LM-7B-alpha Score 56.8%
Zephyr 7B Beta Binary 56%
Starling-LM-7B-alpha Binary 55.7%
Mistral-7B-Instruct-v0.2 Binary 44.7%

Scores at each effort level

Models the leaderboard ran at two or more effort levels. We don’t have enough cost data here to say what the extra effort costs.

Lowest effort to highest, light to dark
Claude 3.5 Sonnet
Binary
90.1%
Llama 3-70B
Score
87.2%
Llama 3.1-8B
Binary Prob.
87%
Qwen2-72B
Prob. Binary
86.4%
Mixtral 8x7B
Binary
83.1%
Mistral-7B-Instruct-v0.2
Binary Score Prob.
82.7%
Llama 3-8B
Score Prob.
82.4%
GPT-4o mini
Score
82%
Zephyr 7B Beta
Binary Prob.
81.6%
GPT-4o
Binary
78.6%
Starling-LM-7B-alpha
Binary Prob.
73.4%

Source: The CASE-Bench paper

ModelEffortHarnessScoreCost per taskOutput tokens
Claude 3.5 Sonnet Score 90.1%
Claude 3.5 Sonnet Binary 88.7%
Llama 3-70B Prob. 87.2%
Llama 3.1-8B Score 87%
Llama 3-70B Binary 86.6%
Qwen2-72B Binary 86.4%
Llama 3.1-8B Prob. 84.9%
Qwen2-72B Score 84.7%
Llama 3-70B Score 84.4%
Mixtral 8x7B Score 83.1%
Mixtral 8x7B Prob. 83.1%
Mistral-7B-Instruct-v0.2 Prob. 82.7%
Llama 3-8B Prob. 82.4%
GPT-4o mini Binary 82%
Mixtral 8x7B Binary 81.9%
Zephyr 7B Beta Prob. 81.6%
Qwen2-72B Prob. 81.6%
GPT-4o mini Score 80%
GPT-4o Aug 2024 Score 78.6%
GPT-4o Aug 2024 Binary 77%
Starling-LM-7B-alpha Prob. 73.4%
Llama 3-8B Binary 70.4%
Llama 3.1-8B Binary 70.3%
Llama 3-8B Score 70%
Mistral-7B-Instruct-v0.2 Score 66.3%
Zephyr 7B Beta Score 56.8%
Starling-LM-7B-alpha Score 56.8%
Zephyr 7B Beta Binary 56%
Starling-LM-7B-alpha Binary 55.7%
Mistral-7B-Instruct-v0.2 Binary 44.7%

Number one over time

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

Anthropic
Benchmark paper Passed 85% in Jan 2025, by Claude 3.5 Sonnet

Source: The CASE-Bench paper

Llama 3-70B on the official leaderboard passed 85% in April 2024

Models released before it came out already scored that high. All the scores we have, one dot each, so a model tested at several settings appears more than once.

Plot with other benchmarks
In the benchmark’s paper Best so far
0% 20% 40% 60% 80% 100% 2024 Near the ceiling (85%) Claude 3.5 Sonnet Mixtral 8x7B 0% 20% 40% 60% 80% 100% 2024 Near the ceiling (85%) Claude 3.5 Sonnet Mixtral 8x7B

Source: Official CASE-Bench leaderboard

No lab model card reports it yet

We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.

More in Safety and alignment

All 636 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%
  • Covert misbehavior (Thinking Machines) In use Agentic tasks that check whether a model misbehaves and conceals it. Scored by the rate of covert misbehavior. Inkling 2 Mini 8.30%
  • Destructive actions (Thinking Machines) In use Agentic tasks that check whether a model takes destructive actions it was not authorized to take. Scored by the rate of destructive actions. Inkling 2 Small 4.70%
  • Misbehavior (Thinking Machines) In use Agentic tasks that check whether a model fabricates results, ignores instructions or hides what it did when a task is hard. Scored by the rate of misbehavior. Inkling 2 Mini 9.80%

Data audit

All 8 checks pass. Every benchmark

Catalog record 2 pass
  • Pass Task access All tasks are public.
  • Pass Confidence Checked against a primary source.
Scores 2 pass 3 notes
  • Pass Independent runs Scores from the benchmark’s paper (14 entries).
  • Pass Models scored 11 models we track, plus 3 other entries on its leaderboard.
  • Note Newest score The newest came 2 years ago, in August 2024, for GPT-4o. That’s expected for a saturated test.
  • Note Frontier models None of the top 10 has a score, which is fine for a retired test. The strongest tested is Claude 3.5 Sonnet (#158).
  • Note Headroom Saturated. The best is 90.1%, past 85% since April 2024.
Leaderboard import 3 pass
  • Pass Importer We read the board on arxiv.org: 30 rows for models we track and 3 other entries at the last read.
  • Pass Last read November 20th, 2026.
  • Pass Read worked The last read worked.
Consistency 1 pass 1 note
  • Note Lab claims No model has both a lab number and an independent run.
  • Pass Known issues No published audit found problems.

How we check