Arena Creative Writing
What it measures and how
Read the creative_writing category in the official text_style_control dataset, using style-controlled Bradley–Terry scores and confidence intervals. Do not substitute overall text or professional-writing scores.
How this evidence is used
Counts toward ranking. Receives 5% from the original 10% human-preference budget; overall text keeps 5%, with no increase in total budget. Shares the evidence family with other Arena tasks.
Benchmark results
Fixed rules select one representative configuration per public model. Anonymous test models are omitted. Original source ranks are retained, with up to 30 models per benchmark.
| Source rank | Source model name | Raw result | Representative configuration |
|---|---|---|---|
| 1 | Claude Opus 5.5 (High)Anthropic | 1,519.1 | High reasoning |
| 2 | Gemini 4 Argon (High)Google | 1,515.6 | High reasoning |
| 3 | Claude Fable 5 (High)Anthropic | 1,501.3 | High reasoning |
| 4 | Claude Opus 4.6 (High)Anthropic | 1,499.7 | High reasoning |
| 5 | Gemini 3.7 Flash (High)Google | 1,491 | High reasoning |
| 6 | Claude Opus 4.7 (High)Anthropic | 1,489 | High reasoning |
| 7 | Gemini 3.8 Flash (High)Google | 1,486.9 | High reasoning |
| 8 | Muse Wonder (Max)Meta | 1,485.1 | Max reasoning |
| 9 | Claude Fable 5.1 (Max)Anthropic | 1,484.7 | Max reasoning |
| 10 | Gemini 3 ProGoogle | 1,484.6 | Source default configuration |
| 12 | Gemini 3.1 Pro PreviewGoogle | 1,481 | Source default configuration |
| 15 | gemini-3.6-flash-highGoogle | 1,471 | High reasoning |
| 16 | Claude Opus 5 (Max)Anthropic | 1,471 | Max reasoning |
| 17 | claude-opus-4-5-20251101-high-32kAnthropic | 1,469.4 | High reasoning · 32k |
| 18 | claude-opus-4-8-highAnthropic | 1,469.1 | High reasoning |
| 19 | qwen3.8-maxAlibaba | 1,469 | Source default configuration |
| 20 | gpt-5.6-sol-xhighOpenAI | 1,468.6 | xHigh reasoning |
| 22 | claude-sonnet-5.5-xhighAnthropic | 1,465.1 | xHigh reasoning |
| 23 | Muse SparkMeta | 1,465 | Source default configuration |
| 24 | gemini-3.5-flash-highGoogle | 1,464.3 | High reasoning |
| 25 | grok-4.20-beta1xAI | 1,463.2 | Source default configuration |
| 27 | Kimi K3 (Max)Moonshot AI | 1,461.3 | Max reasoning |
| 29 | Muse Spark 1.3 (Max)Meta | 1,461.2 | Max reasoning |
| 30 | gpt-6.1-sol-maxOpenAI | 1,458.4 | Max reasoning |
| 31 | gemini-3-flashGoogle | 1,456.8 | Source default configuration |
| 32 | gpt-5.5-instantOpenAI | 1,454.5 | Source default configuration |
| 33 | glm-5.2-maxZ.ai | 1,454.5 | Source default configuration |
| 35 | glm-5.3-maxZ.ai | 1,453.2 | Source default configuration |
| 36 | grok-4.5xAI | 1,451.7 | Source default configuration |
| 37 | gpt-5.5-highOpenAI | 1,451.3 | High reasoning |
Limitations and attribution
Open user prompts are classified automatically; they are not all long-form fiction or Chinese creative writing. Votes include personal preferences. Overlap with Arena overall text and professional writing prevents it from counting as another independent organization.
Data license: CC BY 4.0 · Official Arena Hugging Face leaderboard-dataset; retain attribution and source links and disclose aggregation changes.
Results are published by LMArena. Raw scores and the e/acc consensus index use different scales and cannot be added directly.