ARC (AI2 Challenge)
Grade-school science questions, challenge set. A 2018 benchmark, saturated for current models; kept for history.
As of 19 Sept 2026, DeepSeek V3 leads ARC (AI2 Challenge) on BenchLeader with 95.3%, ahead of Llama 3.1 405B at 95.3%, across 77 model configurations with a published result.
- Published by
- Epoch AI Benchmarking Hub
- Category
- Reasoning
- Index weight
- Reference only
- Models
- 77
- Data as of
- 19 Sept 2026
CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.
What the test looks like
Grade-school science questions, challenge set.
How it is scored
Accuracy, as compiled by Epoch AI from published results.
What to keep in mind
Saturated: frontier models score near the ceiling, so it separates only older or smaller models, and results are largely self-reported by developers.
- 1DeepSeek V395.3%
- 2Llama 3.1 405B95.3%
- 3Qwen2.5 72B94.5%
- 4Deepseek v292.2%
- 5phi-3-medium 14B (medium)91.6%
- 6Phi 3 Small90.7%
- 7GPT 3.5 Turbo 110687.4%
- 8Mixtral 8x7B87.3%
- 9Claude Instant86.3%
- 10Stable Beluga 286.1%
- 11Claude Instant85.7%
- 12davinci-00285.2%
- 13Palm 540B85.2%
- 14Phi 3 Mini84.9%
- 15Qwen-14B84.4%
77 of 77
| # | ||||
|---|---|---|---|---|
| 1 |  DeepSeek V3DeepSeekopen | 95.3% | 44.8 | 2024-12-26 |
| 2 |  Llama 3.1 405BMetaopen | 95.3% | 43.5 | 2024-07-23 |
| 3 |  Qwen2.5 72BAlibabaopen | 94.5% | 42.7 | 2024-09-19 |
| 4 | Deepseek v2DeepSeekopen ↗ | 92.2% | – | 2024-05-07 |
| 5 |  phi-3-medium 14BmediumMicrosoft | 91.6% | – | 2024-04-23 |
| 6 | Phi 3 SmallMicrosoft | 90.7% | – | 2024-04-23 |
| 7 |  GPT 3.5 Turbo 1106OpenAI | 87.4% | 39.2 | 2023-11-06 |
| 8 |  Mixtral 8x7BMistral AIopen | 87.3% | 32.1 | 2023-12-11 |
| 9 |  Claude InstantAnthropic | 86.3% | – | 2023-08-09 |
| 10 |  Stable Beluga 2Stability AI | 86.1% | – | 2023-07-20 |
| 11 | Claude InstantAnthropic | 85.7% | – | – |
| 12 | davinci-002OpenAI | 85.2% | – | 2022-03-15 |
| 13 |  Palm 540BGoogle | 85.2% | – | 2022-04-04 |
| 14 | Phi 3 MiniMicrosoftopen | 84.9% | 36.2 | 2024-04-23 |
| 15 | Qwen-14BAlibaba | 84.4% | – | 2023-09-28 |
| 16 | Llama 3-8BMetaopen | 82.8% | 35.6 | 2024-04-18 |
| 17 |  internlm-20bShanghai AI Lab | 81.7% | – | 2023-09-18 |
| 18 | Mistral 7B v0.3Mistral AIopen | 78.6% | 32.1 | 2023-09-27 |
| 19 | Llama 2-70BMetaopen ↗ | 78.3% | 34.9 | 2023-07-18 |
| 20 | Gemma 7BGoogle | 78.3% | – | 2024-02-21 |
| 21 | Phi-2Microsoft | 75.9% | – | 2023-12-12 |
| 22 | Qwen-7BAlibaba | 75.3% | – | 2023-09-28 |
| 23 | Qwen2.5-Coder 32B InstructAlibabaopen ↗ | 70.5% | 42.6 | 2024-09-18 |
| 24 | LLaMA-65BMetaopen | 69.5% | – | 2023-02-24 |
| 25 | internlm-7bShanghai AI Lab | 69.5% | – | 2023-07-05 |
| 26 | PaLM 2-LGoogle | 69.2% | – | 2023-05-17 |
| 27 |  Falcon-180BTII | 67.8% | – | 2023-09-06 |
| 28 | LLaMA-33BMeta | 67.5% | – | 2023-02-24 |
| 29 | Qwen2.5-Coder-14BAlibaba | 66.0% | – | 2024-09-18 |
| 30 | PaLM 2-MGoogle | 64.9% | – | 2023-05-17 |
| 31 | DeepSeek-Coder-V2 236BDeepSeek | 64.3% | – | 2024-06-17 |
| 32 | Falcon-40BTII | 61.9% | – | 2023-03-15 |
| 33 |  chatglm2-6bZhipu AI | 61.0% | – | 2023-06-24 |
| 34 | Qwen2.5-Coder 7B InstructAlibabaopen ↗ | 60.9% | – | 2024-09-18 |
| 35 | Llama 2-13BMetaopen ↗ | 60.3% | 36.5 | 2023-07-18 |
| 36 | PaLM 2-SGoogle | 59.6% | – | 2023-05-17 |
| 37 | DeepSeek-Coder-V2-Lite-BaseDeepSeek | 57.3% | – | 2024-06-13 |
| 38 |  Yi-9B01.AI | 55.6% | – | 2024-03-01 |
| 39 |  Nemotron-4 15BNVIDIA | 55.5% | – | 2024-02-26 |
| 40 |  INTELLECT-1Prime Intellect | 54.5% | – | 2024-11-29 |
| 41 | Llama 2-34BMeta | 54.5% | – | 2023-07-18 |
| 42 | InstructGPT 175BOpenAI | 53.2% | – | 2022-01-27 |
| 43 | Qwen-1_8BAlibaba | 53.2% | – | 2023-11-30 |
| 44 | Qwen2.5-Coder-3BAlibaba | 52.9% | – | 2024-09-18 |
| 45 | LLaMA-13BMeta | 52.7% | – | 2023-02-24 |
| 46 | PaLM 62BGoogle | 52.5% | – | 2022-04-04 |
| 47 |  MPT-30BMosaicML | 50.6% | – | 2023-06-22 |
| 48 | Yi 6B01.AI | 50.3% | – | 2023-11-02 |
| 49 | Falcon-7BTII | 47.9% | – | 2023-04-24 |
| 50 | LLaMA-7BMeta | 47.6% | – | 2023-02-24 |
| 51 | StarCoder 2 15BNVIDIA | 47.2% | – | 2024-02-20 |
| 52 | Llama 2-7BMetaopen ↗ | 45.9% | 34.2 | 2023-07-18 |
| 53 | Qwen2 5 Coder 1 5BAlibaba | 45.2% | – | 2024-09-18 |
| 54 | Phi-1.5Microsoft | 44.4% | – | 2023-09-11 |
| 55 |  vicuna-13b-v1.1LMSYS | 43.2% | – | 2023-04-12 |
| 56 | MPT-7BMosaicML | 42.6% | – | 2023-05-05 |
| 57 | DeepSeek Coder 33BDeepSeek | 42.2% | – | 2023-11-02 |
| 58 | Gemma 2BGoogle | 42.1% | – | 2024-02-21 |
| 59 |  XGen-7BSalesforce | 41.2% | – | 2023-06-27 |
| 60 | GPT-NeoX-20BEleutherAI | 41.1% | – | 2022-04-07 |
Cite as: BenchLeader, “ARC (AI2 Challenge) leaderboard”, https://www.benchleader.com/benchmarks/arc_ai2, data as of 19 Sept 2026.
ARC (AI2 Challenge): questions
- What does ARC (AI2 Challenge) measure?
- Grade-school science questions, challenge set. Scores are reported in percent of tasks solved; higher is better.
- Which AI model leads ARC (AI2 Challenge)?
- DeepSeek V3 leads ARC (AI2 Challenge) with 95.3% as of 19 Sept 2026, ahead of Llama 3.1 405B at 95.3%.
- How many models have ARC (AI2 Challenge) results?
- 77 model configurations have a ARC (AI2 Challenge) result on BenchLeader, all taken from Epoch AI Benchmarking Hub.
- Who runs ARC (AI2 Challenge) and how often is it updated?
- ARC (AI2 Challenge) is published by Epoch AI Benchmarking Hub. BenchLeader re-reads the published results every morning and records the date each result was published.
- Does ARC (AI2 Challenge) count toward the BenchLeader Index?
- No. ARC (AI2 Challenge) is shown for reference but left out of the composite index.