Arena-Hard v2: leaderboard
Prompt benchmark from LMArena, using automated pairwise judging to separate frontier chat-model quality on open-ended tasks.
Metric: Win Rate (%). Source: github.com. Saturation forecast: Estimated already saturated. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O3 (2025-04-16) | 85.9 | #121 |
| 2 | O4 Mini (2025-04-16) (High) | 79.1 | #179 (O4 Mini (2025-04-16)) |
| 3 | O4 Mini (2025-04-16) | 74.6 | #179 |
| 4 | Gemini 2.5 Flash | 68.6 | #245 |
| 5 | O3 Mini (2025-01-31) (High) | 66.1 | #240 (O3 Mini (2025-01-31)) |
| 6 | O1 (2024-12-17) (High) | 61 | #149 (O1 (2024-12-17)) |
| 7 | Qwen 3 235B A22B | 58.4 | #313 |
| 8 | DeepSeek R1 | 58 | #251 |
| 9 | O1 (2024-12-17) | 55.9 | #149 |
| 10 | GPT-4.1 | 50 | #248 |
| 11 | O3 Mini (2025-01-31) | 50 | #240 |
| 12 | GPT-4.5 (Preview) | 50 | #244 |
| 13 | GPT-4.1 Mini | 46.9 | #355 |
| 14 | Qwen 3 32B | 44.5 | #443 |
| 15 | QwQ-32B | 43.5 | #421 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=arena-hard-v2 · How It Works · Data refreshed daily, snapshot 2026-10-20.