benchmark evidence
Chatbot Arena (LMSYS)
Human-preference Bradley-Terry ratings from LMSYS Chatbot Arena. Community-voted pairwise comparisons.
gameableupstream source ->
winner on Chatbot Arena (LMSYS)
direct benchmark result, not a broad vertical composite | source row dated 2026-08-27
scored on 2026-08-27
latest mapped results | top 20
| # | Model | Score | Evidence | Tested |
|---|---|---|---|---|
| 1 | Anthropic: Claude Opus 4.6 | 99.3 | model-only independent_benchmark | 2026-08-27 |
| 2 | Anthropic: Claude Opus 4.7 | 98.6 | model-only independent_benchmark | 2026-08-27 |
| 3 | OpenAI: GPT-5.5 | 94.2 | model-only independent_benchmark | 2026-08-27 |
| 4 | Anthropic: Claude Sonnet 4.6 | 93.0 | model-only independent_benchmark | 2026-08-27 |
| 5 | Anthropic: Claude Opus 4.5 | 92.4 | model-only independent_benchmark | 2026-08-27 |
| 6 | OpenAI: GPT-5.4 | 91.4 | model-only independent_benchmark | 2026-08-27 |
| 7 | Z.ai: GLM 5 | 89.4 | model-only independent_benchmark | 2026-08-27 |
| 8 | DeepSeek: DeepSeek V4 Pro 0423 | 89.4 | model-only independent_benchmark | 2026-08-27 |
| 9 | Anthropic: Claude Sonnet 4.5 | 88.8 | model-only independent_benchmark | 2026-08-27 |
| 10 | OpenAI: GPT-5.4 Mini | 87.1 | model-only independent_benchmark | 2026-08-27 |
| 11 | Google: Gemini 2.5 Pro | 86.4 | model-only independent_benchmark | 2026-08-27 |
| 12 | OpenAI: GPT-5.1 | 84.7 | model-only independent_benchmark | 2026-08-27 |
| 13 | OpenAI: GPT-5.2 | 83.9 | model-only independent_benchmark | 2026-08-27 |
| 14 | DeepSeek: DeepSeek V4 Flash 0423 | 83.9 | model-only independent_benchmark | 2026-08-27 |
| 15 | OpenAI: o3 | 82.8 | model-only independent_benchmark | 2026-08-27 |
| 16 | OpenAI: GPT-5 | 81.6 | model-only independent_benchmark | 2026-05-19 |
| 17 | DeepSeek: DeepSeek V3.2 | 81.3 | model-only independent_benchmark | 2026-08-27 |
| 18 | MoonshotAI: Kimi K2 Thinking | 79.4 | model-only independent_benchmark | 2026-05-19 |
| 19 | OpenAI: GPT-4.1 | 78.6 | model-only independent_benchmark | 2026-08-27 |
| 20 | Anthropic: Claude Haiku 4.5 | 78.3 | model-only independent_benchmark | 2026-08-27 |
what this result means
Human-preference Bradley-Terry ratings from LMSYS Chatbot Arena. Community-voted pairwise comparisons.
Human preference is useful, but presentation style and familiarity can move arena votes.
A win here is a win on Chatbot Arena (LMSYS). Broad task pages require independent corroboration before naming a general winner.
source record
category: overall
metric: bradley_terry
matched models: 34
latest source date: 2026-08-27
direction: higher is better