ArxivRollBench
Models reconstruct passages from recently published scientific papers. They order fragments, fill gaps and choose what comes next.
- Made by
- Zi Liang and colleagues
- Released
- July 2025
- Kind
- Reading comprehension
- Level
- University
- Scored
- Checked automatically
- Answers
- Multiple choice
- Tasks
- Public, plus a private test set
- Measures
- Capability
- Language
- English
- Interaction
- Single turn
- Items from
- Real work and use
- Refresh
- New items on a schedule
- Run by
- A host
- In views
- Kept fresh
- Strongest model tested
- GPT-5.5 #21 on the capability index
- Last model added
- May 2026
- Scores refreshed
- November 20th, 2026
Kimi K2.6 leads with 70.8% among models released by May 2026 Kimi K2.6 leads with 70.8% among models released by May 2026
Best score per entry from the benchmark's own leaderboard, read on November 20th, 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. This view shows only models we track, each at its best.
The newest model with a score here came out in May 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 16 more
- 70.8 Kimi K2.6
- 55.2
Gemini 3.5 Flash
- 54.4 GPT-5.5
- 53.4 Claude Opus 4.7
- 53.3
DeepSeek-V4-Pro
- 52.0 Grok 4.3
- 48.3 Claude Sonnet 4.6
- 46.0 Kimi K2.5
- 45.3
DeepSeek-V4-Flash
- 43.2
DeepSeek-R1
- 42.9 Claude Opus 4.6
- 42.1 GPT-5.3 Codex
- 40.8 Claude Haiku 4.5
- 39.3 GPT-5.4
- 38.5
Qwen3-14B
- 37.6 deepseek/deepseek-chat-v3-0324
- 35.5 GPT-5.4 Mini
- 34.5 Grok 4.20
- 33.9
DeepSeek-V3.1
- 31.9
DeepSeek-V3.2
- 30.0 Qwen/Qwen2.5-7B-Instruct
- 29.5
Qwen2-7B-Instruct
- 29.3
Qwen3-8B
- 28.2
Qwen2.5-7B
- 27.7 meta-llama/Llama-3.1-8B-Instruct
- 27.4
Qwen3-4B
- 26.2 mistralai/Mistral-7B-Instruct-v0.3
- 25.8
Mistral-7B-Instruct-v0.1
- 23.8
Falcon3-10B-Instruct
- 23.7
Mistral-7B-Instruct-v0.2
- 23.4
Llama 3.2 3B
- 23.2 meta-llama/Llama-3.2-3B-Instruct
- 22.1
Llama 3.2 1B
- 21.9
Llama 3.1-8B
- 21.6 Qwen/Qwen2.5-Math-7B
- 21.2
Phi-4-Reasoning-plus
- 12.6 meta-llama/Llama-3.2-1B-Instruct
- 8.25
Phi-4-reasoning
- 7.55 Qwen/Qwen2.5-Math-7B-Instruct
- 1 Kimi K2.6 2026a, 50-item task samples 70.8%
- 2
Gemini 3.5 Flash 2026a, 50-item task samples 55.2%
- 3 GPT-5.5 2026a, 50-item task samples 54.4%
- 4 Claude Opus 4.7 2026a, 50-item task samples 53.4%
- 5
DeepSeek-V4-Pro 2026a, 50-item task samples 53.3%
- 6 Grok 4.3 2026a, 50-item task samples 52%
- 7 Claude Sonnet 4.6 2026a, 50-item task samples 48.3%
- 8 Kimi K2.5 2026a, 50-item task samples 46%
- 9
DeepSeek-V4-Flash 2026a, 50-item task samples 45.3%
- 10
DeepSeek-R1 May 2025, 2026a, 50-item task samples 43.2%
- 11 Claude Opus 4.6 2026a, 50-item task samples 42.9%
- 12 GPT-5.3 Codex 2026a, 50-item task samples 42.1%
- 13 Claude Haiku 4.5 2026a, 50-item task samples 40.8%
- 14 GPT-5.4 2026a, 50-item task samples 39.3%
- 15
Qwen3-14B 2026a, full tasks 38.5%
- 16 deepseek/deepseek-chat-v3-0324 2026a, 50-item task samples 37.6%
- 17 GPT-5.4 Mini 2026a, 50-item task samples 35.5%
- 18 Grok 4.20 2026a, 50-item task samples 34.5%
- 19
DeepSeek-V3.1 2026a, 50-item task samples 33.9%
- 20
DeepSeek-V3.2 2026a, 50-item task samples 31.9%
- 21 Qwen/Qwen2.5-7B-Instruct 2026a, full tasks 30%
- 22
Qwen2-7B-Instruct 2026a, full tasks 29.5%
- 23
Qwen3-8B 2026a, full tasks 29.3%
- 24
Qwen2.5-7B 2026a, full tasks 28.2%
- 25 meta-llama/Llama-3.1-8B-Instruct 2026a, full tasks 27.7%
- 26
Qwen3-4B 2026a, full tasks 27.4%
- 27 mistralai/Mistral-7B-Instruct-v0.3 2026a, full tasks 26.2%
- 28
Mistral-7B-Instruct-v0.1 2026a, full tasks 25.8%
- 29
Falcon3-10B-Instruct 2026a, full tasks 23.8%
- 30
Mistral-7B-Instruct-v0.2 2026a, full tasks 23.7%
- 31
Llama 3.2 3B 2026a, full tasks 23.4%
- 32 meta-llama/Llama-3.2-3B-Instruct 2026a, full tasks 23.2%
- 33
Llama 3.2 1B 2026a, full tasks 22.1%
- 34
Llama 3.1-8B 2026a, full tasks 21.9%
- 35 Qwen/Qwen2.5-Math-7B 2026a, full tasks 21.6%
- 36
Phi-4-Reasoning-plus 2026a, full tasks 21.2%
- 37 meta-llama/Llama-3.2-1B-Instruct 2026a, full tasks 12.6%
- 38
Phi-4-reasoning 2026a, full tasks 8.25%
- 39 Qwen/Qwen2.5-Math-7B-Instruct 2026a, full tasks 7.55%
Source: Official ArxivRollBench leaderboard
- 70.8 Kimi K2.6
- 55.2
Gemini 3.5 Flash
- 54.4 GPT-5.5
- 53.4 Claude Opus 4.7
- 53.3
DeepSeek-V4-Pro
- 52.0 Grok 4.3
- 48.3 Claude Sonnet 4.6
- 46.0 Kimi K2.5
- 45.3
DeepSeek-V4-Flash
- 43.2
DeepSeek-R1
- 42.9 Claude Opus 4.6
- 42.1 GPT-5.3 Codex
- 40.8 Claude Haiku 4.5
- 39.3 GPT-5.4
- 38.5
Qwen3-14B
- 35.5 GPT-5.4 Mini
- 34.5 Grok 4.20
- 33.9
DeepSeek-V3.1
- 31.9
DeepSeek-V3.2
- 29.5
Qwen2-7B-Instruct
- 29.3
Qwen3-8B
- 28.2
Qwen2.5-7B
- 27.4
Qwen3-4B
- 25.8
Mistral-7B-Instruct-v0.1
- 23.8
Falcon3-10B-Instruct
- 23.7
Mistral-7B-Instruct-v0.2
- 23.4
Llama 3.2 3B
- 22.1
Llama 3.2 1B
- 21.9
Llama 3.1-8B
- 21.2
Phi-4-Reasoning-plus
- 8.25
Phi-4-reasoning
- 1 Kimi K2.6 2026a, 50-item task samples 70.8%
- 2
Gemini 3.5 Flash 2026a, 50-item task samples 55.2%
- 3 GPT-5.5 2026a, 50-item task samples 54.4%
- 4 Claude Opus 4.7 2026a, 50-item task samples 53.4%
- 5
DeepSeek-V4-Pro 2026a, 50-item task samples 53.3%
- 6 Grok 4.3 2026a, 50-item task samples 52%
- 7 Claude Sonnet 4.6 2026a, 50-item task samples 48.3%
- 8 Kimi K2.5 2026a, 50-item task samples 46%
- 9
DeepSeek-V4-Flash 2026a, 50-item task samples 45.3%
- 10
DeepSeek-R1 May 2025, 2026a, 50-item task samples 43.2%
- 11 Claude Opus 4.6 2026a, 50-item task samples 42.9%
- 12 GPT-5.3 Codex 2026a, 50-item task samples 42.1%
- 13 Claude Haiku 4.5 2026a, 50-item task samples 40.8%
- 14 GPT-5.4 2026a, 50-item task samples 39.3%
- 15
Qwen3-14B 2026a, full tasks 38.5%
- 16 GPT-5.4 Mini 2026a, 50-item task samples 35.5%
- 17 Grok 4.20 2026a, 50-item task samples 34.5%
- 18
DeepSeek-V3.1 2026a, 50-item task samples 33.9%
- 19
DeepSeek-V3.2 2026a, 50-item task samples 31.9%
- 20
Qwen2-7B-Instruct 2026a, full tasks 29.5%
- 21
Qwen3-8B 2026a, full tasks 29.3%
- 22
Qwen2.5-7B 2026a, full tasks 28.2%
- 23
Qwen3-4B 2026a, full tasks 27.4%
- 24
Mistral-7B-Instruct-v0.1 2026a, full tasks 25.8%
- 25
Falcon3-10B-Instruct 2026a, full tasks 23.8%
- 26
Mistral-7B-Instruct-v0.2 2026a, full tasks 23.7%
- 27
Llama 3.2 3B 2026a, full tasks 23.4%
- 28
Llama 3.2 1B 2026a, full tasks 22.1%
- 29
Llama 3.1-8B 2026a, full tasks 21.9%
- 30
Phi-4-Reasoning-plus 2026a, full tasks 21.2%
- 31
Phi-4-reasoning 2026a, full tasks 8.25%
Source: Official ArxivRollBench leaderboard
| # | Entry | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Kimi K2.6 | 70.8% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 2 | | 55.2% | 2026a, 50-item task samples | May 2026 | Official leaderboard |
| 3 | GPT-5.5 | 54.4% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 4 | Claude Opus 4.7 | 53.4% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 5 | | 53.3% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 6 | Grok 4.3 | 52% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 7 | Claude Sonnet 4.6 | 48.3% | 2026a, 50-item task samples | Feb 2026 | Official leaderboard |
| 8 | Kimi K2.5 | 46% | 2026a, 50-item task samples | Jan 2026 | Official leaderboard |
| 9 | | 45.3% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 10 | | 43.2% | 2026a, 50-item task samples | May 2025 | Official leaderboard |
| 11 | Claude Opus 4.6 | 42.9% | 2026a, 50-item task samples | Feb 2026 | Official leaderboard |
| 12 | GPT-5.3 Codex | 42.1% | 2026a, 50-item task samples | Feb 2026 | Official leaderboard |
| 13 | Claude Haiku 4.5 | 40.8% | 2026a, 50-item task samples | Oct 2025 | Official leaderboard |
| 14 | GPT-5.4 | 39.3% | 2026a, 50-item task samples | Mar 2026 | Official leaderboard |
| 15 | | 38.5% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 16 | deepseek/deepseek-chat-v3-0324 | 37.6% | 2026a, 50-item task samples | Official leaderboard | |
| 17 | GPT-5.4 Mini | 35.5% | 2026a, 50-item task samples | Mar 2026 | Official leaderboard |
| 18 | Grok 4.20 | 34.5% | 2026a, 50-item task samples | Feb 2026 | Official leaderboard |
| 19 | | 33.9% | 2026a, 50-item task samples | Aug 2025 | Official leaderboard |
| 20 | | 31.9% | 2026a, 50-item task samples | Dec 2025 | Official leaderboard |
| 21 | Qwen/Qwen2.5-7B-Instruct | 30% | 2026a, full tasks | Official leaderboard | |
| 22 | | 29.5% | 2026a, full tasks | Jun 2024 | Official leaderboard |
| 23 | | 29.3% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 24 | | 28.2% | 2026a, full tasks | Sep 2024 | Official leaderboard |
| 25 | meta-llama/Llama-3.1-8B-Instruct | 27.7% | 2026a, full tasks | Official leaderboard | |
| 26 | | 27.4% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 27 | mistralai/Mistral-7B-Instruct-v0.3 | 26.2% | 2026a, full tasks | Official leaderboard | |
| 28 | | 25.8% | 2026a, full tasks | Sep 2023 | Official leaderboard |
| 29 | | 23.8% | 2026a, full tasks | Dec 2024 | Official leaderboard |
| 30 | | 23.7% | 2026a, full tasks | Dec 2023 | Official leaderboard |
| 31 | | 23.4% | 2026a, full tasks | Sep 2024 | Official leaderboard |
| 32 | meta-llama/Llama-3.2-3B-Instruct | 23.2% | 2026a, full tasks | Official leaderboard | |
| 33 | | 22.1% | 2026a, full tasks | Sep 2024 | Official leaderboard |
| 34 | | 21.9% | 2026a, full tasks | Jul 2024 | Official leaderboard |
| 35 | Qwen/Qwen2.5-Math-7B | 21.6% | 2026a, full tasks | Official leaderboard | |
| 36 | | 21.2% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 37 | meta-llama/Llama-3.2-1B-Instruct | 12.6% | 2026a, full tasks | Official leaderboard | |
| 38 | | 8.25% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 39 | Qwen/Qwen2.5-Math-7B-Instruct | 7.55% | 2026a, full tasks | Official leaderboard |
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | Kimi K2.6 | 70.8% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 2 | | 55.2% | 2026a, 50-item task samples | May 2026 | Official leaderboard |
| 3 | GPT-5.5 | 54.4% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 4 | Claude Opus 4.7 | 53.4% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 5 | | 53.3% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 6 | Grok 4.3 | 52% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 7 | Claude Sonnet 4.6 | 48.3% | 2026a, 50-item task samples | Feb 2026 | Official leaderboard |
| 8 | Kimi K2.5 | 46% | 2026a, 50-item task samples | Jan 2026 | Official leaderboard |
| 9 | | 45.3% | 2026a, 50-item task samples | Apr 2026 | Official leaderboard |
| 10 | | 43.2% | 2026a, 50-item task samples | May 2025 | Official leaderboard |
| 11 | Claude Opus 4.6 | 42.9% | 2026a, 50-item task samples | Feb 2026 | Official leaderboard |
| 12 | GPT-5.3 Codex | 42.1% | 2026a, 50-item task samples | Feb 2026 | Official leaderboard |
| 13 | Claude Haiku 4.5 | 40.8% | 2026a, 50-item task samples | Oct 2025 | Official leaderboard |
| 14 | GPT-5.4 | 39.3% | 2026a, 50-item task samples | Mar 2026 | Official leaderboard |
| 15 | | 38.5% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 16 | GPT-5.4 Mini | 35.5% | 2026a, 50-item task samples | Mar 2026 | Official leaderboard |
| 17 | Grok 4.20 | 34.5% | 2026a, 50-item task samples | Feb 2026 | Official leaderboard |
| 18 | | 33.9% | 2026a, 50-item task samples | Aug 2025 | Official leaderboard |
| 19 | | 31.9% | 2026a, 50-item task samples | Dec 2025 | Official leaderboard |
| 20 | | 29.5% | 2026a, full tasks | Jun 2024 | Official leaderboard |
| 21 | | 29.3% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 22 | | 28.2% | 2026a, full tasks | Sep 2024 | Official leaderboard |
| 23 | | 27.4% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 24 | | 25.8% | 2026a, full tasks | Sep 2023 | Official leaderboard |
| 25 | | 23.8% | 2026a, full tasks | Dec 2024 | Official leaderboard |
| 26 | | 23.7% | 2026a, full tasks | Dec 2023 | Official leaderboard |
| 27 | | 23.4% | 2026a, full tasks | Sep 2024 | Official leaderboard |
| 28 | | 22.1% | 2026a, full tasks | Sep 2024 | Official leaderboard |
| 29 | | 21.9% | 2026a, full tasks | Jul 2024 | Official leaderboard |
| 30 | | 21.2% | 2026a, full tasks | Apr 2025 | Official leaderboard |
| 31 | | 8.25% | 2026a, full tasks | Apr 2025 | Official leaderboard |
Rankings by part
We read 24 parts from the benchmark’s own leaderboard, each ranking models on one part of the test, such as cs: Cloze, cs: Prediction and cs: Sequencing. Kimi K2.6 leads 16 of them, and 4 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
cs: Cloze
31 models we track have a score here.
- 1 Kimi K2.6 50%
- 2
Mistral-7B-Instruct-v0.1 ↑22 28.3%
- 3
Qwen2-7B-Instruct ↑17 25.4%
- 4
Qwen2.5-7B ↑18 25.3%
- 5
DeepSeek-R1 May 2025 ↑5 25%
- 6
DeepSeek-V4-Flash ↑3 25%
- 7
Qwen3-4B ↑16 24.6%
- 8
Qwen3-14B ↑7 24.2%
- 9
Llama 3.2 3B ↑18 24.1%
- 10
Qwen3-8B ↑11 24%
cs: Prediction
31 models we track have a score here.
- 1 Kimi K2.6 100%
- 2
Gemini 3.5 Flash 84.3%
- 3 GPT-5.5 72.5%
- 4 Kimi K2.5 ↑4 66.7%
- 5 Claude Sonnet 4.6 ↑2 62.7%
- 6 Claude Opus 4.6 ↑5 60%
- 7 Grok 4.3 ↓1 60%
- 8 Claude Opus 4.7 ↓4 58.8%
- 9 GPT-5.3 Codex ↑3 56.3%
- 10
DeepSeek-V4-Pro ↓5 55.2%
cs: Sequencing
30 models we track have a score here.
- 1
DeepSeek-R1 May 2025 ↑9 63.6%
- 2 Grok 4.3 ↑4 58%
- 3
Gemini 3.5 Flash ↓1 56.9%
- 4 GPT-5.5 ↓1 54.9%
- 5
DeepSeek-V4-Flash ↑4 54.1%
- 6 Claude Sonnet 4.6 ↑1 52.9%
- 7 Claude Opus 4.7 ↓3 52.9%
- 8
DeepSeek-V4-Pro ↓3 51.4%
- 9 Claude Haiku 4.5 ↑4 48%
- 10 GPT-5.3 Codex ↑2 47.1%
econ: Cloze
30 models we track have a score here.
- 1 Kimi K2.5 ↑7 35.7%
- 2 GPT-5.5 ↑1 34%
- 3 Grok 4.3 ↑3 34%
- 4 Claude Opus 4.6 ↑7 31.9%
- 5
DeepSeek-R1 May 2025 ↑5 30.8%
- 6 Claude Opus 4.7 ↓2 29.8%
- 7
Gemini 3.5 Flash ↓5 29.8%
- 8
DeepSeek-V4-Pro ↓3 27.8%
- 9
Qwen2.5-7B ↑13 27.7%
- 10
Qwen3-8B ↑11 26.7%
econ: Prediction
31 models we track have a score here.
- 1 Kimi K2.6 83.3%
- 2 GPT-5.5 ↑1 78.4%
- 3 Claude Opus 4.6 ↑8 75%
- 4 Claude Opus 4.7 66.7%
- 5
Gemini 3.5 Flash ↓3 66.7%
- 6 Grok 4.3 56.9%
- 7 Claude Sonnet 4.6 54.9%
- 8 GPT-5.3 Codex ↑4 52%
- 9
DeepSeek-V4-Pro ↓4 48%
- 10
Qwen3-14B ↑5 45.1%
econ: Sequencing
31 models we track have a score here.
- 1 Kimi K2.6 100%
- 2 Kimi K2.5 ↑6 73.3%
- 3 Claude Opus 4.7 ↑1 70.8%
- 4
DeepSeek-V4-Pro ↑1 66.7%
- 5
DeepSeek-V4-Flash ↑4 63.6%
- 6 Claude Sonnet 4.6 ↑1 62.5%
- 7
Gemini 3.5 Flash ↓5 62.5%
- 8 Grok 4.3 ↓2 60.4%
- 9 GPT-5.5 ↓6 58.3%
- 10
DeepSeek-R1 May 2025 57.1%
eess: Cloze
31 models we track have a score here.
- 1 Kimi K2.6 50%
- 2
DeepSeek-V4-Flash ↑7 33.3%
- 3
DeepSeek-R1 May 2025 ↑7 29.4%
- 4
Mistral-7B-Instruct-v0.1 ↑20 27.1%
- 5 Claude Haiku 4.5 ↑8 26.7%
- 6 Claude Opus 4.7 ↓2 26.7%
- 7 GPT-5.5 ↓4 26.7%
- 8 Grok 4.3 ↓2 26.7%
- 9
Qwen2.5-7B ↑13 26.1%
- 10
Qwen2-7B-Instruct ↑10 25.4%
eess: Prediction
30 models we track have a score here.
- 1 Kimi K2.6 80%
- 2 GPT-5.5 ↑1 78.4%
- 3 Grok 4.3 ↑3 76.5%
- 4
Gemini 3.5 Flash ↓2 74.5%
- 5 Claude Opus 4.7 ↓1 72%
- 6 Kimi K2.5 ↑2 71.4%
- 7 Claude Sonnet 4.6 68.6%
- 8
DeepSeek-V4-Pro ↓3 67.7%
- 9 GPT-5.3 Codex ↑3 66.7%
- 10 Grok 4.20 ↑7 62.7%
eess: Sequencing
31 models we track have a score here.
- 1 Kimi K2.6 57.1%
- 2
DeepSeek-R1 May 2025 ↑8 55.6%
- 3
DeepSeek-V4-Pro ↑2 55.6%
- 4 Claude Opus 4.7 54.9%
- 5 GPT-5.5 ↓2 54.9%
- 6 GPT-5.4 ↑8 51%
- 7 Grok 4.3 ↓1 51%
- 8
DeepSeek-V3.1 ↑10 49%
- 9 Claude Opus 4.6 ↑2 49%
- 10 GPT-5.4 Mini ↑6 49%
math: Cloze
31 models we track have a score here.
- 1
DeepSeek-V4-Pro ↑4 38.5%
- 2 Claude Opus 4.6 ↑9 33.3%
- 3
DeepSeek-R1 May 2025 ↑7 30%
- 4 Claude Opus 4.7 30%
- 5 GPT-5.5 ↓2 30%
- 6 Grok 4.3 30%
- 7
Gemini 3.5 Flash ↓5 30%
- 8
Qwen2.5-7B ↑14 29.8%
- 9
Qwen3-14B ↑6 28.7%
- 10
Mistral-7B-Instruct-v0.1 ↑14 27.6%
math: Prediction
30 models we track have a score here.
- 1 Kimi K2.6 83.3%
- 2
Gemini 3.5 Flash 80.4%
- 3
DeepSeek-V4-Pro ↑2 76.9%
- 4 Claude Sonnet 4.6 ↑3 76.5%
- 5 Claude Opus 4.7 ↓1 72.5%
- 6 GPT-5.5 ↓3 72.5%
- 7
DeepSeek-R1 May 2025 ↑3 70%
- 8 Kimi K2.5 69.2%
- 9 GPT-5.3 Codex ↑3 68%
- 10 Grok 4.3 ↓4 68%
math: Sequencing
31 models we track have a score here.
- 1 Kimi K2.5 ↑7 76.9%
- 2 Kimi K2.6 ↓1 66.7%
- 3 Claude Opus 4.7 ↑1 65.9%
- 4
Gemini 3.5 Flash ↓2 65.9%
- 5
DeepSeek-V4-Pro 64.7%
- 6
DeepSeek-V4-Flash ↑3 64%
- 7 Claude Opus 4.6 ↑4 63.6%
- 8
DeepSeek-R1 May 2025 ↑2 62.5%
- 9 GPT-5.4 Mini ↑7 61.4%
- 10 GPT-5.5 ↓7 59.1%
physics: Cloze
31 models we track have a score here.
- 1 Kimi K2.6 50%
- 2 Grok 4.3 ↑4 34.1%
- 3
Gemini 3.5 Flash ↓1 34.1%
- 4 Kimi K2.5 ↑4 33.3%
- 5 GPT-5.5 ↓2 31.7%
- 6 Claude Opus 4.7 ↓2 30%
- 7 Claude Opus 4.6 ↑4 29.3%
- 8 Claude Haiku 4.5 ↑5 26.8%
- 9
Mistral-7B-Instruct-v0.1 ↑15 26.7%
- 10
Qwen2-7B-Instruct ↑10 25.9%
physics: Prediction
30 models we track have a score here.
- 1 Claude Opus 4.7 ↑3 67.3%
- 2 Kimi K2.6 ↓1 66.7%
- 3 Claude Sonnet 4.6 ↑4 66%
- 4 GPT-5.5 ↓1 64%
- 5
DeepSeek-V4-Pro 61.9%
- 6
Gemini 3.5 Flash ↓4 60%
- 7 GPT-5.3 Codex ↑5 58.3%
- 8 GPT-5.4 ↑6 56%
- 9 Grok 4.3 ↓3 54%
- 10 Claude Haiku 4.5 ↑3 50%
physics: Sequencing
31 models we track have a score here.
- 1 Kimi K2.6 75%
- 2 Claude Opus 4.6 ↑9 55.3%
- 3 Claude Opus 4.7 ↑1 55.3%
- 4 GPT-5.5 ↓1 55.3%
- 5 Kimi K2.5 ↑3 54.5%
- 6
Gemini 3.5 Flash ↓4 51.1%
- 7
DeepSeek-V4-Pro ↓2 46.2%
- 8 Grok 4.3 ↓2 44.7%
- 9
DeepSeek-V3.2 ↑10 38.3%
- 10 Claude Sonnet 4.6 ↓3 38.3%
q-bio: Cloze
31 models we track have a score here.
- 1 Kimi K2.5 ↑7 45.5%
- 2 GPT-5.5 ↑1 40%
- 3
Gemini 3.5 Flash ↓1 40%
- 4 Claude Opus 4.7 37.8%
- 5 Grok 4.3 ↑1 37.5%
- 6 Claude Opus 4.6 ↑5 35%
- 7 Claude Sonnet 4.6 35%
- 8
DeepSeek-R1 May 2025 ↑2 33.3%
- 9 Claude Haiku 4.5 ↑4 30%
- 10
Mistral-7B-Instruct-v0.1 ↑14 28.8%
q-bio: Prediction
30 models we track have a score here.
- 1 Kimi K2.6 80%
- 2 GPT-5.5 ↑1 77.1%
- 3
Gemini 3.5 Flash ↓1 75%
- 4 Kimi K2.5 ↑4 72.7%
- 5 GPT-5.3 Codex ↑7 66.7%
- 6 Claude Sonnet 4.6 ↑1 66.7%
- 7
DeepSeek-V4-Flash ↑2 64.7%
- 8 GPT-5.4 ↑6 64.6%
- 9 Grok 4.3 ↓3 64.6%
- 10 Claude Opus 4.7 ↓6 64.4%
q-bio: Sequencing
31 models we track have a score here.
- 1 Kimi K2.6 100%
- 2 GPT-5.3 Codex ↑10 78.1%
- 3 Claude Opus 4.7 ↑1 76.1%
- 4
Gemini 3.5 Flash ↓2 74.5%
- 5 Grok 4.3 ↑1 72.3%
- 6
DeepSeek-V4-Pro ↓1 72%
- 7 Claude Opus 4.6 ↑4 70.2%
- 8 Kimi K2.5 66.7%
- 9 GPT-5.5 ↓6 63.8%
- 10 Claude Sonnet 4.6 ↓3 61.7%
q-fin: Cloze
31 models we track have a score here.
- 1 Kimi K2.6 66.7%
- 2
DeepSeek-V4-Pro ↑3 37.5%
- 3
Mistral-7B-Instruct-v0.1 ↑21 29.5%
- 4
Qwen2.5-7B ↑18 26.7%
- 5
Qwen2-7B-Instruct ↑15 25.7%
- 6
Qwen3-14B ↑9 24.6%
- 7 Claude Opus 4.6 ↑4 24.4%
- 8 Claude Opus 4.7 ↓4 24.4%
- 9
Gemini 3.5 Flash ↓7 24.4%
- 10
Llama 3.2 3B ↑17 22.9%
q-fin: Prediction
30 models we track have a score here.
- 1 Kimi K2.6 100%
- 2
DeepSeek-V4-Pro ↑3 83.3%
- 3 GPT-5.5 82.4%
- 4
DeepSeek-R1 May 2025 ↑6 80%
- 5 Claude Sonnet 4.6 ↑2 76.5%
- 6 Grok 4.3 76.5%
- 7
Gemini 3.5 Flash ↓5 74.5%
- 8 Claude Opus 4.7 ↓4 72.5%
- 9 GPT-5.3 Codex ↑3 69.4%
- 10 Claude Haiku 4.5 ↑3 64.7%
q-fin: Sequencing
31 models we track have a score here.
- 1
DeepSeek-V4-Pro ↑4 83.3%
- 2
DeepSeek-V4-Flash ↑7 70.4%
- 3 Grok 4.3 ↑3 66.7%
- 4 Claude Opus 4.6 ↑7 60.4%
- 5
Gemini 3.5 Flash ↓3 60.4%
- 6
DeepSeek-R1 May 2025 ↑4 58.3%
- 7 Claude Opus 4.7 ↓3 58.3%
- 8 GPT-5.5 ↓5 56.3%
- 9 Claude Sonnet 4.6 ↓2 54.2%
- 10 Claude Haiku 4.5 ↑3 52.1%
stat: Cloze
31 models we track have a score here.
- 1 Kimi K2.6 100%
- 2
DeepSeek-V4-Pro ↑3 42.1%
- 3 Kimi K2.5 ↑5 41.7%
- 4
DeepSeek-V4-Flash ↑5 37.5%
- 5 Claude Opus 4.6 ↑6 28.3%
- 6 GPT-5.5 ↓3 28.3%
- 7 Grok 4.3 ↓1 28.3%
- 8
Gemini 3.5 Flash ↓6 28.3%
- 9
Mistral-7B-Instruct-v0.1 ↑15 27.6%
- 10
Qwen2.5-7B ↑12 27.3%
stat: Prediction
30 models we track have a score here.
- 1 Kimi K2.6 100%
- 2
Gemini 3.5 Flash 74.5%
- 3
DeepSeek-V4-Pro ↑2 70.8%
- 4 Claude Opus 4.7 68.6%
- 5 GPT-5.5 ↓2 62.7%
- 6 Grok 4.3 62.7%
- 7 Kimi K2.5 ↑1 61.5%
- 8 GPT-5.3 Codex ↑4 61.2%
- 9 Claude Sonnet 4.6 ↓2 60.8%
- 10 GPT-5.4 ↑4 58.8%
stat: Sequencing
31 models we track have a score here.
- 1
DeepSeek-V4-Pro ↑4 54.5%
- 2 Grok 4.3 ↑4 53.1%
- 3
Gemini 3.5 Flash ↓1 53.1%
- 4 Kimi K2.5 ↑4 50%
- 5 Claude Opus 4.7 ↓1 50%
- 6 Kimi K2.6 ↓5 50%
- 7 Claude Opus 4.6 ↑4 46.9%
- 8 GPT-5.5 ↓5 46.9%
- 9
DeepSeek-V3.2 ↑10 44.9%
- 10 GPT-5.4 ↑4 44.9%
5 models have held number one on the official leaderboard since its release in July 2025
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official ArxivRollBench leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official ArxivRollBench 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 Knowledge
- MMLU-Pro In use Harder 10-option version of MMLU with reasoning-focused questions across 14 subjects.
Intern-S2-397B 89.8%
- MMLU In use 57-subject multiple-choice exam questions from elementary to professional level. Scored by accuracy.
Seed1.8 92.3%
- SuperGPQA In use 26,529 graduate-level multiple-choice questions across 285 disciplines. Accuracy.
Qwen 3.6 Max (Preview) 73.9%
- WideSearch In use 200 bilingual broad info-seeking tasks: agent collects many atomic facts into a table. Scored by success rate and row/item F1.
Tencent Hy4 preview 83.9%
- BBEH micro average In use Models solve harder replacements for each of the 23 BIG-Bench Hard tasks, 4,520 examples in total. The score is the share of all examples answered correctly.
Gemma 4 31B IT 74.4%
- HLE-Diamond In use Models answer 1,000 cleaned expert questions from Humanity's Last Exam, 500 on reasoning and 500 on knowledge, closed book. Graded on exact answers. Claude Fable 5.2 77.9%
Data audit
9 of 10 checks pass, 1 warning. Look first at the Newest score check. The newest came 6 months ago, in May 2026, for Gemini 3.5 Flash. Every benchmark
Catalog record 2 pass
- Pass Task access Public tasks, plus a private test set.
- Pass Confidence Checked against a primary source.
Scores 1 warning 3 pass 1 note
- Pass Independent runs Scores from the official leaderboard (39 entries).
- Pass Models scored 31 models we track, plus 8 other entries on its leaderboard.
- Warning Newest score The newest came 6 months ago, in May 2026, for Gemini 3.5 Flash.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.5 (#21).
- Pass Headroom The best is 70.8%, under 85%, so it still separates models.
Leaderboard import 3 pass
- Pass Importer We read the board on arxivroll.moreoverai.com: 31 rows for models we track and 8 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.