reasoning benchmark
Arena-Hard v2 Leaderboard
Arena-Hard v2 is a benchmark of 500 prompts from Chatbot Arena and WildChat-1M for evaluating large language models on real-world user queries across software engineering, mathematics, creative writing, and technical problem-solving.
Updated Sep 5, 2026
Models19
Model coverage19
MetricScore
EvidenceB
Arena-Hard v2 Ranking
Higher score ranks better on this benchmark.
Sort by
Rank | Model | Score | Percentile | Participants | Evidence | Evaluated |
|---|
| Rank01 | ModelXI | Score86.20% | Percentile100.00% | Participants19 | EvidenceC | Evaluated |
| Rank02 | ModelAC | Score82.70% | Percentile94.44% | Participants19 | EvidenceC | Evaluated |
| Rank03 | ModelAC | Score79.70% | Percentile88.89% | Participants19 | EvidenceC | Evaluated |
| Rank04 | ModelAC | Score79.20% | Percentile83.33% | Participants19 | EvidenceC | Evaluated |
| Rank05 | ModelAC | Score77.40% | Percentile77.78% | Participants19 | EvidenceC | Evaluated |
| Rank06 | ModelNV | Score73.88% | Percentile72.22% | Participants19 | EvidenceC | Evaluated |
| Rank07 | ModelSA | Score71.00% | Percentile66.67% | Participants19 | EvidenceC | Evaluated |
| Rank08 | ModelIB | Score67.93% | Percentile61.11% | Participants19 | EvidenceC | Evaluated |
| Rank09 | ModelNV | Score67.70% | Percentile55.56% | Participants19 | EvidenceC | Evaluated |
| Rank10 | ModelIB | Score65.19% | Percentile50.00% | Participants19 | EvidenceC | Evaluated |
| Rank11 | ModelAC | Score64.70% | Percentile44.44% | Participants19 | EvidenceC | Evaluated |
| Rank12 | ModelAC | Score62.30% | Percentile38.89% | Participants19 | EvidenceC | Evaluated |
| Rank13 | ModelAC | Score60.50% | Percentile33.33% | Participants19 | EvidenceC | Evaluated |
| Rank14 | ModelAC | Score58.50% | Percentile27.78% | Participants19 | EvidenceC | Evaluated |
| Rank15 | ModelAC | Score56.70% | Percentile22.22% | Participants19 | EvidenceC | Evaluated |
| Rank16 | ModelAC | Score51.10% | Percentile16.67% | Participants19 | EvidenceC | Evaluated |
| Rank17 | ModelSA | Score49.00% | Percentile11.11% | Participants19 | EvidenceC | Evaluated |
| Rank18 | ModelAC | Score36.80% | Percentile5.56% | Participants19 | EvidenceC | Evaluated |
| Rank19 | ModelIB | Score34.96% | Percentile0.00% | Participants19 | EvidenceC | Evaluated |
Arena-Hard v2 Highlights
The leading models and scores on this benchmark.
Arena-Hard v2 Score Distribution
A closer view of the leading scores on this benchmark.
The Top AI Models for Arena-Hard v2
The first five results on this benchmark, with official price and output speed added where the model identity can be matched.
What is Arena-Hard v2?
What Arena-Hard v2 measures and how its scores work.
Arena-Hard v2 is a benchmark consisting of 500 curated prompts sourced from Chatbot Arena and WildChat-1M, with automatic evaluation using LLM-as-a-Judge.
It measures large language model performance on open-ended software engineering problems, mathematics, creative writing, technical problem-solving, and other real-world user queries using the Score metric as a ratio.
Scores are shown in ratio. This benchmark is not independently verified and has an evidence level of B.
- Family
- Arena-Hard v2
- Modality
- text
- Primary category
- reasoning
- Score direction
- higher
- LLMBoard eligible
- No
- Evaluation key
- overall
LLM Stats. Benchmark scores retain their original unit. Overall score eligibility is shown separately.
FAQ
Common questions about Arena-Hard v2.
Which model scores highest on Arena-Hard v2?
MiMo-V2-Flash is currently ranked first with 86.20%.
What are the top three models on Arena-Hard v2?
The current leaders are MiMo-V2-Flash (86.20%), Qwen3-Next-80B-A3B-Instruct (82.70%), and Qwen3-235B-A22B-Thinking-2507 (79.70%).
Which Arena-Hard v2 model has the lowest official input price?
MiMo-V2-Flash has the lowest matched official input price at $0.14 input / $0.28 output per 1M tokens.
Which models are fastest among Arena-Hard v2 results?
The fastest matched records are Qwen3 VL 4B Thinking (8.77 tok/s via DeepInfra).
Does the highest score result prove overall model quality?
No. This benchmark measures one defined capability or task. The overall LLMBoard score uses a separate aggregation across eligible benchmark evidence.