chat benchmark
Arena Hard is an automatic evaluation benchmark for instruction-tuned LLMs based on 500 challenging real-world prompts curated by BenchBuilder.
Updated Sep 5, 2026
Higher score ranks better on this benchmark.
Rank | Model | Score | Percentile | Participants | Evidence | Evaluated |
|---|
| Rank01 | ModelAC | Score95.60% | Percentile100.00% | Participants26 | EvidenceC | Evaluated |
| Rank02 | ModelAC | Score93.80% | Percentile96.00% | Participants26 | EvidenceC | Evaluated |
| Rank03 | ModelAC | Score91.00% | Percentile92.00% | Participants26 | EvidenceC | Evaluated |
| Rank04 | ModelNV | Score88.30% | Percentile88.00% | Participants26 | EvidenceC | Evaluated |
| Rank05 | ModelMA | Score87.60% | Percentile84.00% | Participants26 | EvidenceC | Evaluated |
| Rank06 | ModelAC | Score81.20% | Percentile80.00% | Participants26 | EvidenceC | Evaluated |
| Rank07 | ModelMI | Score79.00% | Percentile76.00% | Participants26 | EvidenceC | Evaluated |
| Rank08 | ModelDE | Score76.20% | Percentile72.00% | Participants26 | EvidenceC | Evaluated |
| Rank09 | ModelMI | Score75.40% | Percentile68.00% | Participants26 | EvidenceC | Evaluated |
| Rank10 | ModelMI | Score73.30% | Percentile64.00% | Participants26 | EvidenceC | Evaluated |
| Rank11 | ModelMA | Score70.90% | Percentile60.00% | Participants26 | EvidenceC | Evaluated |
| Rank12 | ModelAL | Score65.40% | Percentile56.00% | Participants26 | EvidenceC | Evaluated |
| Rank13 | ModelMA | Score58.30% | Percentile52.00% | Participants26 | EvidenceC | Evaluated |
| Rank14 | ModelIB | Score57.56% | Percentile48.00% | Participants26 | EvidenceC | Evaluated |
| Rank15 | ModelIB | Score57.56% | Percentile44.00% | Participants26 | EvidenceC | Evaluated |
| Rank16 | ModelMA | Score55.10% | Percentile40.00% | Participants26 | EvidenceC | Evaluated |
| Rank17 | ModelMA | Score55.10% | Percentile36.00% | Participants26 | EvidenceC | Evaluated |
| Rank18 | ModelAC | Score52.00% | Percentile32.00% | Participants26 | EvidenceC | Evaluated |
| Rank19 | ModelMA | Score50.90% | Percentile28.00% | Participants26 | EvidenceC | Evaluated |
| Rank20 | ModelAL | Score46.10% | Percentile24.00% | Participants26 | EvidenceC | Evaluated |
| Rank21 | ModelMA | Score43.10% | Percentile20.00% | Participants26 | EvidenceC | Evaluated |
| Rank22 | ModelMI | Score37.90% | Percentile16.00% | Participants26 | EvidenceC | Evaluated |
| Rank23 | ModelMI | Score37.00% | Percentile12.00% | Participants26 | EvidenceC | Evaluated |
| Rank24 | ModelMI | Score32.80% | Percentile8.00% | Participants26 | EvidenceC | Evaluated |
| Rank25 | ModelMA | Score30.50% | Percentile4.00% | Participants26 | EvidenceC | Evaluated |
| Rank26 | ModelIB | Score26.70% | Percentile0.00% | Participants26 | EvidenceC | Evaluated |
The leading models and scores on this benchmark.
A closer view of the leading scores on this benchmark.
The first five results on this benchmark, with official price and output speed added where the model identity can be matched.
What Arena Hard measures and how its scores work.
Arena-Hard-Auto is a benchmark comprising open-ended software engineering problems, mathematical questions, and creative writing tasks, evaluated with GPT-4.1 and Gemini-2.5 as automatic judges.
It measures model performance using LLM-as-a-Judge evaluations that approximate human preference, reported as a Score with unit ratio.
Scores are shown in ratio. This benchmark is not independently verified and has an evidence level of B.
LLM Stats. Benchmark scores retain their original unit. Overall score eligibility is shown separately.
Common questions about Arena Hard.
Qwen3 235B A22B is currently ranked first with 95.60%.
The current leaders are Qwen3 235B A22B (95.60%), Qwen3 32B (93.80%), and Qwen3 30B A3B (91.00%).
Phi 4 Mini has the lowest matched official input price at $0.08 input / $0.30 output per 1M tokens.
The fastest matched records are Mistral Small 4 (163.71 tok/s via Mistral AI), Mistral Small 3 24B Instruct (134.00 tok/s via Mistral AI), and DeepSeek-V2.5 (100.00 tok/s via DeepSeek).
No. This benchmark measures one defined capability or task. The overall LLMBoard score uses a separate aggregation across eligible benchmark evidence.
It measures model performance using LLM-as-a-Judge evaluations that approximate human preference, reported as a Score with unit ratio.
Yes. Higher values rank better for this benchmark.
26 model results are currently shown.
No. This benchmark is shown for reference but does not contribute to the overall score.
Ranking basisThis arena hard AI model leaderboard uses descending score in the benchmark original unit. The leaderboard ranking keeps matched price and speed data separate from benchmark evidence.
Selection summary
Qwen3 235B A22B currently leads Arena Hard with 95.60%. 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.