reasoning benchmark
Arena Hard Leaderboard
Arena-Hard-Auto is an automatic evaluation benchmark for instruction-tuned LLMs consisting of 500 challenging real-world prompts curated by BenchBuilder. It includes open-ended software engineering problems, mathematical questions, and creative writing tasks. The benchmark uses LLM-as-a-Judge methodology with GPT-4.1 and Gemini-2.5 as automatic judges to approximate human preference. Arena-Hard achieves 98.6% correlation with human preference rankings and provides 3x higher separation of model performances compared to MT-Bench, making it highly effective for distinguishing between models of similar quality.
Updated Aug 17, 2026
Models26
Model coverage26
MetricScore
EvidenceB
Arena Hard Ranking
Higher score ranks better on this benchmark.
Sort by
Rank | Model | Score | Percentile | Participants | Evidence | Evaluated |
|---|
| Rank01 | ModelAC | Score95.6% | Percentile100.0% | Participants26 | EvidenceC | Evaluated |
| Rank02 | ModelAC | Score93.8% | Percentile96.0% | Participants26 | EvidenceC | Evaluated |
| Rank03 | ModelAC | Score91.0% | Percentile92.0% | Participants26 | EvidenceC | Evaluated |
| Rank04 | ModelNV | Score88.3% | Percentile88.0% | Participants26 | EvidenceC | Evaluated |
| Rank05 | ModelMA | Score87.6% | Percentile84.0% | Participants26 | EvidenceC | Evaluated |
| Rank06 | ModelAC | Score81.2% | Percentile80.0% | Participants26 | EvidenceC | Evaluated |
| Rank07 | ModelMI | Score79.0% | Percentile76.0% | Participants26 | EvidenceC | Evaluated |
| Rank08 | ModelDE | Score76.2% | Percentile72.0% | Participants26 | EvidenceC | Evaluated |
| Rank09 | ModelMI | Score75.4% | Percentile68.0% | Participants26 | EvidenceC | Evaluated |
| Rank10 | ModelMI | Score73.3% | Percentile64.0% | Participants26 | EvidenceC | Evaluated |
| Rank11 | ModelMA | Score70.9% | Percentile60.0% | Participants26 | EvidenceC | Evaluated |
| Rank12 | ModelAL | Score65.4% | Percentile56.0% | Participants26 | EvidenceC | Evaluated |
| Rank13 | ModelMA | Score58.3% | Percentile52.0% | Participants26 | EvidenceC | Evaluated |
| Rank14 | ModelIB | Score57.6% | Percentile48.0% | Participants26 | EvidenceC | Evaluated |
| Rank15 | ModelIB | Score57.6% | Percentile44.0% | Participants26 | EvidenceC | Evaluated |
| Rank16 | ModelMA | Score55.1% | Percentile40.0% | Participants26 | EvidenceC | Evaluated |
| Rank17 | ModelMA | Score55.1% | Percentile36.0% | Participants26 | EvidenceC | Evaluated |
| Rank18 | ModelAC | Score52.0% | Percentile32.0% | Participants26 | EvidenceC | Evaluated |
| Rank19 | ModelMA | Score50.9% | Percentile28.0% | Participants26 | EvidenceC | Evaluated |
| Rank20 | ModelAL | Score46.1% | Percentile24.0% | Participants26 | EvidenceC | Evaluated |
| Rank21 | ModelMA | Score43.1% | Percentile20.0% | Participants26 | EvidenceC | Evaluated |
| Rank22 | ModelMI | Score37.9% | Percentile16.0% | Participants26 | EvidenceC | Evaluated |
| Rank23 | ModelMI | Score37.0% | Percentile12.0% | Participants26 | EvidenceC | Evaluated |
| Rank24 | ModelMI | Score32.8% | Percentile8.0% | Participants26 | EvidenceC | Evaluated |
| Rank25 | ModelMA | Score30.5% | Percentile4.0% | Participants26 | EvidenceC | Evaluated |
| Rank26 | ModelIB | Score26.7% | Percentile0.0% | Participants26 | EvidenceC | Evaluated |
Arena Hard Highlights
The leading models and scores on this benchmark.
Arena Hard Score Distribution
A closer view of the leading scores on this benchmark.
The Top AI Models for Arena Hard
The first five results on this benchmark, with official price and output speed added where the model identity can be matched.
Ranking basisThis arena hard AI model leaderboard uses descending score in the benchmark's original unit. The leaderboard ranking keeps matched price and speed data separate from benchmark evidence.
- 01ACQwen3 235B A22BAlibaba Cloud / Qwen Team
- Score
- 95.6%
- Price
- $0.70 input / $2.8 output per 1M tokens
- Speed
- Up to 68 tok/s via Fireworks
Strengths
- Ranks #1 of 26 compared models
- 100th percentile on this benchmark
- C evidence result
Considerations
- This result measures Arena Hard, not total model capability
- 02ACQwen3 32BAlibaba Cloud / Qwen Team
- Score
- 93.8%
- Price
- $0.70 input / $2.8 output per 1M tokens
- Speed
- Up to 328 tok/s via Sambanova
Strengths
- Ranks #2 of 26 compared models
- 96th percentile on this benchmark
- C evidence result
Considerations
- This result measures Arena Hard, not total model capability
- 03ACQwen3 30B A3BAlibaba Cloud / Qwen Team
- Score
- 91.0%
- Speed
- Up to 122 tok/s via Fireworks
Strengths
- Ranks #3 of 26 compared models
- 92th percentile on this benchmark
- C evidence result
Considerations
- This result measures Arena Hard, not total model capability
- 04NV
- Score
- 88.3%
Strengths
- Ranks #4 of 26 compared models
- 88th percentile on this benchmark
- C evidence result
Considerations
- This result measures Arena Hard, not total model capability
- 05MAMistral Small 3 24B InstructMistral AI
- Score
- 87.6%
- Speed
- Up to 134 tok/s via Mistral AI
Strengths
- Ranks #5 of 26 compared models
- 84th percentile on this benchmark
- C evidence result
Considerations
- This result measures Arena Hard, not total model capability
Selection summary
Best AI Models for Arena Hard
Qwen3 235B A22B currently leads Arena Hard with 95.6%. It is the top model on this specific benchmark, while the best LLM for the broader task should also be checked against other benchmarks, price and runtime.
Use this leaderboard with the supporting benchmark results and coverage details above. A leaderboard position summarizes the selected ranking signal; it does not replace workload-specific testing.
What is Arena Hard?
What Arena Hard measures and how its scores work.
Arena-Hard-Auto is an automatic evaluation benchmark for instruction-tuned LLMs consisting of 500 challenging real-world prompts curated by BenchBuilder. It includes open-ended software engineering problems, mathematical questions, and creative writing tasks. The benchmark uses LLM-as-a-Judge methodology with GPT-4.1 and Gemini-2.5 as automatic judges to approximate human preference. Arena-Hard achieves 98.6% correlation with human preference rankings and provides 3x higher separation of model performances compared to MT-Bench, making it highly effective for distinguishing between models of similar quality.
Scores are shown in ratio. This benchmark is not independently verified and has an evidence level of B.
- Family
- Arena Hard
- Modality
- text
- Primary category
- reasoning
- Score direction
- higher
- LLMBoard eligible
- Yes
- Evaluation key
- arena-hard|llm-stats-current
Benchmark scores retain their original unit. Overall score eligibility is shown separately.
FAQ
Common questions about Arena Hard.
Which model scores highest on Arena Hard?
Qwen3 235B A22B is currently ranked first with 95.6%.
What does Arena Hard measure?
Arena-Hard-Auto is an automatic evaluation benchmark for instruction-tuned LLMs consisting of 500 challenging real-world prompts curated by BenchBuilder. It includes open-ended software engineering problems, mathematical questions, and creative writing tasks. The benchmark uses LLM-as-a-Judge methodology with GPT-4.1 and Gemini-2.5 as automatic judges to approximate human preference. Arena-Hard achieves 98.6% correlation with human preference rankings and provides 3x higher separation of model performances compared to MT-Bench, making it highly effective for distinguishing between models of similar quality.
Is a higher score better?
Yes. Higher values rank better for this benchmark.
How many models are compared?
26 model results are currently shown.
Does this benchmark affect the overall score?
Yes. This benchmark can contribute to the current LLMBoard capability score.