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.
MiMo-V2-Flash from Xiaomi currently leads the Arena-Hard v2 leaderboard with a score of 0.862 across 16 evaluated AI models.
What Arena-Hard v2 measures
Arena-Hard v2 is a text benchmark that evaluates large language models on reasoning, general, writing, and creativity tasks. LLM Stats tracks 16 models on this benchmark, with a maximum possible score of 1. Current average across reported models is 0.7, with the leader reaching 0.9.
Compare leaders on the best AI for reasoning, best AI for general, best AI for writing and best AI for creativity leaderboards.
Publication
- Paper
- From Crowdsourced Data to High-Quality Benchmarks: Arena-Hard and BenchBuilder Pipeline
- Authors
- Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, and 4 others
- Published
- arXiv
- 2406.11939
Abstract
The rapid evolution of Large Language Models (LLMs) has outpaced the development of model evaluation, highlighting the need for continuous curation of new, challenging benchmarks. However, manual curation of high-quality, human-aligned benchmarks is expensive and time-consuming. To address this, we introduce BenchBuilder, an automated pipeline that leverages LLMs to curate high-quality, open-ended prompts from large, crowd-sourced datasets, enabling continuous benchmark updates without human in the loop. We apply BenchBuilder to datasets such as Chatbot Arena and WildChat-1M, extracting challenging prompts and utilizing LLM-as-a-Judge for automatic model evaluation. To validate benchmark quality, we propose new metrics to measure a benchmark's alignment with human preferences and ability to separate models. We release Arena-Hard-Auto, a benchmark consisting 500 challenging prompts curated by BenchBuilder. Arena-Hard-Auto provides 3x higher separation of model performances compared to MT-Bench and achieves 98.6% correlation with human preference rankings, all at a cost of $20. Our work sets a new framework for the scalable curation of automated benchmarks from extensive data.
MiMo-V2-Flash leads with 86.2%, followed by 
Qwen3-Next-80B-A3B-Instruct at 82.7% and Qwen3-235B-A22B-Thinking-2507 at 79.7%.
Progress Over Time
Interactive timeline showing model performance evolution on Arena-Hard v2
Arena-Hard v2 Leaderboard
| Context | Cost | License | ||||
|---|---|---|---|---|---|---|
| 1 | Xiaomi | 309B | — | — | ||
| 2 |  Alibaba Cloud / Qwen Team | 80B | — | — | ||
| 3 | Alibaba Cloud / Qwen Team | 235B | — | — | ||
| 4 | Alibaba Cloud / Qwen Team | 235B | — | — | ||
| 5 | Alibaba Cloud / Qwen Team | 236B | 262K | $0.30 / $1.50 | ||
| 6 | 120B | — | — | |||
| 7 |  Sarvam AI | 105B | — | — | ||
| 8 | 32B | 262K | $0.06 / $0.24 | |||
| 9 | Alibaba Cloud / Qwen Team | 33B | — | — | ||
| 10 | Alibaba Cloud / Qwen Team | 80B | — | — | ||
| 11 | Alibaba Cloud / Qwen Team | 33B | — | — | ||
| 12 | Alibaba Cloud / Qwen Team | 31B | — | — | ||
| 13 | Alibaba Cloud / Qwen Team | 31B | — | — | ||
| 14 | Alibaba Cloud / Qwen Team | 9B | 262K | $0.18 / $2.09 | ||
| 15 | Sarvam AI | 30B | — | — | ||
| 16 | Alibaba Cloud / Qwen Team | 4B | 262K | $0.10 / $1.00 |
FAQ
Common questions about Arena-Hard v2.
More evaluations to explore
Related benchmarks in the same category
A challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. Questions are Google-proof and extremely difficult, with PhD experts reaching 65% accuracy.
A more robust and challenging multi-task language understanding benchmark that extends MMLU by expanding multiple-choice options from 4 to 10, eliminating trivial questions, and focusing on reasoning-intensive tasks. Features over 12,000 curated questions across 14 domains and causes a 16-33% accuracy drop compared to original MMLU.
All 30 problems from the 2025 American Invitational Mathematics Examination (AIME I and AIME II), testing olympiad-level mathematical reasoning with integer answers from 000-999. Used as an AI benchmark to evaluate large language models' ability to solve complex mathematical problems requiring multi-step logical deductions and structured symbolic reasoning.
Massive Multitask Language Understanding benchmark testing knowledge across 57 diverse subjects including STEM, humanities, social sciences, and professional domains
A verified subset of 500 software engineering problems from real GitHub issues, validated by human annotators for evaluating language models' ability to resolve real-world coding issues by generating patches for Python codebases.
Humanity's Last Exam (HLE) is a multi-modal academic benchmark with 2,500 questions across mathematics, humanities, and natural sciences, designed to test LLM capabilities at the frontier of human knowledge with unambiguous, verifiable solutions