reasoning benchmark
Arena-Hard v2
Arena-Hard-Auto v2 is a challenging benchmark consisting of 500 carefully curated prompts sourced from Chatbot Arena and WildChat-1M, designed to evaluate large language models on real-world user queries. The benchmark covers diverse domains including open-ended software engineering problems, mathematics, creative writing, and technical problem-solving. It uses LLM-as-a-Judge for automatic evaluation, achieving 98.6% correlation with human preference rankings while providing 3x higher separation of model performances compared to MT-Bench. The benchmark emphasizes prompt specificity, complexity, and domain knowledge to better distinguish between model capabilities.
Updated Aug 7, 2026
Published models16
Registry coverage16
MetricScore
EvidenceB
Arena-Hard v2 leaderboard
Sorted by the source-provided rank. Higher score is better according to the registry.
| 1 | XI | 86.2% | 100.0% | 16 | C | |
| 2 | AC | 82.7% | 93.3% | 16 | C | |
| 3 | AC | 79.7% | 86.7% | 16 | C | |
| 4 | AC | 79.2% | 80.0% | 16 | C | |
| 5 | AC | 77.4% | 73.3% | 16 | C | |
| 6 | NV | 73.9% | 66.7% | 16 | C | |
| 7 | SA | 71.0% | 60.0% | 16 | C | |
| 8 | NV | 67.7% | 53.3% | 16 | C | |
| 9 | AC | 64.7% | 46.7% | 16 | C | |
| 10 | AC | 62.3% | 40.0% | 16 | C | |
| 11 | AC | 60.5% | 33.3% | 16 | C | |
| 12 | AC | 58.5% | 26.7% | 16 | C | |
| 13 | AC | 56.7% | 20.0% | 16 | C | |
| 14 | AC | 51.1% | 13.3% | 16 | C | |
| 15 | SA | 49.0% | 6.7% | 16 | C | |
| 16 | AC | 36.8% | 0.0% | 16 | C |
Score distribution
Top published rows on the benchmark's original scale.
Arena-Hard v2 highlights
The top published results on this benchmark's own scale.
What is Arena-Hard v2?
Definition and scoring fields from the benchmark registry.
Arena-Hard-Auto v2 is a challenging benchmark consisting of 500 carefully curated prompts sourced from Chatbot Arena and WildChat-1M, designed to evaluate large language models on real-world user queries. The benchmark covers diverse domains including open-ended software engineering problems, mathematics, creative writing, and technical problem-solving. It uses LLM-as-a-Judge for automatic evaluation, achieving 98.6% correlation with human preference rankings while providing 3x higher separation of model performances compared to MT-Bench. The benchmark emphasizes prompt specificity, complexity, and domain knowledge to better distinguish between model capabilities.
Scores are shown in ratio. The current registry marks this benchmark as not independently verified with evidence level B.
- Family
- Arena-Hard v2
- Modality
- text
- Primary category
- reasoning
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- arena-hard-v2|llm-stats-current
Source-native results are preserved. Eligibility for the overall LLMBoard score is a separate policy decision.
FAQ
Common questions about Arena-Hard v2.
Which model scores highest on Arena-Hard v2?
MiMo-V2-Flash is currently ranked first with 86.2%.
What does Arena-Hard v2 measure?
Arena-Hard-Auto v2 is a challenging benchmark consisting of 500 carefully curated prompts sourced from Chatbot Arena and WildChat-1M, designed to evaluate large language models on real-world user queries. The benchmark covers diverse domains including open-ended software engineering problems, mathematics, creative writing, and technical problem-solving. It uses LLM-as-a-Judge for automatic evaluation, achieving 98.6% correlation with human preference rankings while providing 3x higher separation of model performances compared to MT-Bench. The benchmark emphasizes prompt specificity, complexity, and domain knowledge to better distinguish between model capabilities.
Is a higher score better?
Yes. Higher values rank better for this benchmark.
How many models are compared?
16 unique published model results are currently shown.
Does this benchmark affect the overall score?
This benchmark is preserved as source-native evidence but is not eligible for the current overall score.