🔌
New: connect Claude & other AIs to GenAIList over MCP — research the catalog and contribute to the shared knowledge base. Learn how →
// BENCHMARK
ArenaHard benchmark
AI model leaderboard for the ArenaHard benchmark. Compare how large language models score on ArenaHard, see the full ranking, and understand what this AI benchmark measures. Qwen3-235B-A22B currently leads with 96.1%. Crowdsourced hard-prompt Arena set graded by GPT-4 judge.
Leaderboard
| # | Model | Organization | Score | Variant | Source |
|---|---|---|---|---|---|
| #1 | Qwen3-235B-A22B | Qwen | 96.1% | non-thinking | official ↗ |
| #2 | Qwen3-235B-A22B | Qwen | 95.6% | thinking | official ↗ |
| #3 | Qwen3-32B | Qwen | 93.8% | thinking | official ↗ |
| #4 | Pangu Pro MoE | Huawei | 93.6% | instruct | official ↗ |
| #5 | Qwen3-32B | Qwen | 92.8% | non-thinking | official ↗ |
| #6 | GPT-4o (Nov 2024) | OpenAI | 92.4% | cited-minimax-text | official ↗ |
| #7 | DeepSeek-V3 | DeepSeek | 92.1% | official ↗ | |
| #8 | Llama Nemotron Super 49B | NVIDIA | 92% | reasoning-on-pass@1 | official ↗ |
| #9 | Qwen3-14B | Qwen | 91.7% | thinking | official ↗ |
| #10 | DeepSeek-V3 | DeepSeek | 91.4% | cited-minimax-text | official ↗ |
| #11 | Qwen3-30B-A3B | Qwen | 91% | thinking | official ↗ |
| #12 | MiniMax-Text-01 | MiniMax | 89.1% | from-hf-readme | official ↗ |
| #13 | Qwen3-30B-A3B | Qwen | 88% | non-thinking | official ↗ |
| #14 | Claude 3.5 Sonnet | Anthropic | 87.6% | 1022-cited-minimax-text | official ↗ |
| #15 | dots.llm1 | Rednote | 87.1% | GPT-4o judge | official ↗ |
| #16 | Llama 4 Scout | Meta AI | 86.9% | cited-pangu-pro-moe | official ↗ |
| #17 | Qwen3-14B | Qwen | 86.3% | non-thinking | official ↗ |
| #18 | Llama-Primus-Nemotron-70B | Trend Micro | 85.8% | from-hf-readme | official ↗ |
| #19 | Qwen3-8B | Qwen | 85.8% | thinking | official ↗ |
| #20 | Gemini 1.5 Pro | Google DeepMind | 85.3% | 002-cited-minimax-text | official ↗ |
| #21 | Qwen2.5 Instruct (72B) | Qwen | 81.2% | instruct-cited-minimax-text | official ↗ |
| #22 | Qwen3-8B | Qwen | 79.6% | non-thinking | official ↗ |
| #23 | Claude 3.5 Sonnet | Anthropic | 79.3% | cited-athene | official ↗ |
| #24 | GPT-4o (May 2024) | OpenAI | 79.2% | cited-athene | official ↗ |
| #25 | EXAONE 3.5 32B | LG AI Research | 78.6% | instruct-from-hf-readme | official ↗ |
| #26 | Athene-70B | Nexusflow | 77.8% | from-hf-readme | official ↗ |
| #27 | Qwen3-4B | Qwen | 76.6% | thinking | official ↗ |
| #28 | DeepSeek-V2.5 | DeepSeek | 76.2% | from-hf-readme | official ↗ |
| #29 | GPT-4o mini | OpenAI | 75% | cited-phi35moe | official ↗ |
| #30 | Gemini 2.0 Flash | Google DeepMind | 72.7% | exp-cited-minimax-text | official ↗ |
| #31 | Gemini 1.5 Pro | Google DeepMind | 72% | cited-athene | official ↗ |
| #32 | EXAONE 3.5 7.8B | LG AI Research | 68.7% | instruct-from-hf-readme | official ↗ |
| #33 | Qwen2.5-32B | Qwen | 67% | instruct-cited-exaone35 | official ↗ |
| #34 | Qwen3-4B | Qwen | 66.2% | non-thinking | official ↗ |
| #35 | Llama 3.1-405B | Meta AI | 63.5% | instruct-cited-minimax-text | official ↗ |
| #36 | Gemma 2 27B | Google DeepMind | 57.5% | instruct-cited-exaone35 | official ↗ |
| #37 | Gemma 2 27B | Google DeepMind | 57% | cited-athene | official ↗ |
| #38 | Qwen2.5-7B | Qwen | 55.5% | cited-phi4-mini | official ↗ |
| #39 | Granite 3.2 8B | IBM | 55.25% | instruct-from-hf-readme | official ↗ |
| #40 | Gemini 1.5 Flash (Sep 2024) | Google DeepMind | 55.2% | cited-phi35moe | official ↗ |
| #41 | Nemotron-4 340B | NVIDIA | 54.2% | instruct-from-hf-readme | official ↗ |
| #42 | Qwen2.5-7B | Qwen | 48.9% | instruct-cited-exaone35-7b | official ↗ |
| #43 | EXAONE 3.5 2.4B | LG AI Research | 48.2% | instruct-from-hf-readme | official ↗ |
| #44 | Llama 3-70B | Meta AI | 46.6% | cited-athene | official ↗ |
| #45 | Granite-4.0-H-Small | IBM | 46.48% | official ↗ | |
| #46 | Gemma 2 9B | Google DeepMind | 43.6% | instruct-cited-exaone35 | official ↗ |
| #47 | Qwen3-1.7B | Qwen | 43.1% | thinking | official ↗ |
| #48 | Gemma 2 9B | Google DeepMind | 42% | instruct-cited-phi35moe | official ↗ |
| #49 | Mistral NeMo | Mistral AI | 39.4% | instruct-cited-phi35moe | official ↗ |
| #50 | Phi-3.5-MoE | Microsoft | 37.9% | instruct-from-hf-readme | official ↗ |
| #51 | Granite 3.1 8B | IBM | 37.58% | instruct-cited-granite-3-2 | official ↗ |
| #52 | Qwen3-1.7B | Qwen | 36.9% | non-thinking | official ↗ |
| #53 | Llama 3.1-8B | Meta AI | 36.43% | instruct-cited-granite-3-2 | official ↗ |
| #54 | Granite-4.0-H-Micro | IBM | 36.15% | official ↗ | |
| #55 | Granite-4.0-H-Tiny | IBM | 35.75% | official ↗ | |
| #56 | phi-3.5-mini | Microsoft | 34.4% | cited-phi4-mini | official ↗ |
| #57 | Phi-4 Mini | Microsoft | 32.8% | instruct-from-hf-readme | official ↗ |
| #58 | Qwen2.5-3B | Qwen | 32% | cited-phi4-mini | official ↗ |
| #59 | Llama 3.1-8B | Meta AI | 27.7% | instruct-cited-exaone35 | official ↗ |
| #60 | Ministral 8B | Mistral AI | 26.9% | cited-phi4-mini | official ↗ |
| #61 | Qwen2.5-3B | Qwen | 26.4% | instruct-cited-exaone35 | official ↗ |
| #62 | Llama 3.1-8B | Meta AI | 25.7% | instruct-cited-phi35moe | official ↗ |
| #63 | Qwen2.5-7B | Qwen | 25.44% | instruct-cited-granite-3-2 | official ↗ |
| #64 | Granite 3.2 2B | IBM | 24.86% | instruct-from-hf-readme | official ↗ |
| #65 | Granite 3.1 2B | IBM | 23.3% | instruct-cited-granite-3-2 | official ↗ |
| #66 | Yi-1.5-34B | 01.AI | 23.1% | instruct-cited-exaone35 | official ↗ |
| #67 | Gemma 2 2B | Google DeepMind | 19.1% | instruct-cited-exaone35 | official ↗ |
| #68 | DeepSeek-R1-Distill-Llama-8B | DeepSeek | 17.17% | cited-granite-3-2 | official ↗ |
| #69 | Llama 3.2 3B | Meta AI | 17% | cited-phi4-mini | official ↗ |
| #70 | Llama 3.2 3B | Meta AI | 14.2% | instruct-cited-exaone35 | official ↗ |
| #71 | Qwen2.5-1.5B | Qwen | 10.6% | instruct-cited-exaone35 | official ↗ |
| #72 | DeepSeek-R1-Distill-Qwen-7B | DeepSeek | 10.36% | cited-granite-3-2 | official ↗ |
| #73 | DeepSeek-R1-Distill-Qwen-7B | Qwen | 10.36% | cited-granite-3-2 | official ↗ |
| #74 | Qwen3-0.6B | Qwen | 8.5% | thinking | official ↗ |
| #75 | Qwen3-0.6B | Qwen | 6.5% | non-thinking | official ↗ |
Frequently asked questions about ArenaHard
What is the ArenaHard benchmark?
Crowdsourced hard-prompt Arena set graded by GPT-4 judge.
How is the ArenaHard benchmark scored?
ArenaHard is scored using the win-rate (%) metric, where a higher score is better. The maximum achievable score is 100.000. GenAIList aggregates reported scores from model providers and papers into a single ranked leaderboard.
Which AI model scores highest on ArenaHard?
As of the latest reported scores on GenAIList, Qwen3-235B-A22B achieves the highest result on ArenaHard with a score of 96.1%.
Is a higher ArenaHard score better?
Yes. On ArenaHard a higher score indicates better performance, so models near the top of the leaderboard are the strongest.