Ru Arena Hard: leaderboard
Russian-language Arena Hard Auto: pairwise comparison of LLM responses to Russian prompts using GPT-4 as judge, reporting win rates.
Metric: Win Rate (%). Source: huggingface.co. Status: saturated. 137 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | DeepSeek V3 Chat | 96.3 |
| 2 | GPT-4o ChatGPT | 94.75 |
| 3 | O1 Mini | 93.45 |
| 4 | Claude 3 Opus (20240229) | 91.3 |
| 5 | GPT-4 Preview (1106) | 90.89 |
| 6 | O1 Preview | 90.8 |
| 7 | Gemini 1.5 Pro (002) | 89.08 |
| 8 | Qwen 2.5 72B Instruct | 88.25 |
| 9 | Claude 3.5 Sonnet (20240620) | 88.17 |
| 10 | Gemma 2 27B (IT) | 86.87 |
| 11 | Phi-4 | 86.59 |
| 12 | Claude 3 Sonnet (20240229) | 84.27 |
| 13 | GPT-4o Mini | 83.9 |
| 14 | Llama 3.1 70B Instruct | 83.26 |
| 15 | Claude 3 Haiku (20240307) | 80.74 |
Interactive version: theaggregate.ai/benchmark?slug=ru-arena-hard · How It Works · Data refreshed daily, snapshot 2026-09-05.