SimAgentsWriting
Can a model tell a good story?
Models write stories from the same briefs. Other AI models compare the stories in both reading orders and judge which is better.
Writing score
Scores reflect AI judges’ preferences; zero is the group average. Dots show scores. Lines show 95% uncertainty ranges: smaller differences may not be reliable.
All results
Download table ↓| Rank | Model and settings | 95% uncertainty range | Details | ||
|---|---|---|---|---|---|
| 1 | Claude Opus 5 (xhigh) | 3.777 | 3.722 to 3.839 | ||
| 2 | Claude Fable 5.1 (high) | 3.760 | 3.675 to 3.860 | ||
| 3 | Claude Opus 5.5 (high) | 3.718 | 3.630 to 3.810 | ||
| 4 | GPT-6 Astra (high) | 3.461 | 3.388 to 3.535 | ||
| 5 | Claude Opus 5 (high) | 3.414 | 3.313 to 3.531 | ||
| 6 | GLM-5.3 (max) | 3.025 | 2.928 to 3.113 | ||
| 7 | Claude Fable 5 (high)Stories completed: 395/400 | 2.901 | 2.838 to 2.974 | ||
| 8 | GPT-5.6 Sol (xhigh) | 2.519 | 2.452 to 2.600 | ||
| 9 | GPT-5.5 (xhigh) | 2.482 | 2.415 to 2.565 | ||
| 10 | Kimi K3 | 2.469 | 2.387 to 2.546 | ||
| 11 | GPT-5.6 Sol (high) | 2.329 | 2.250 to 2.409 | ||
| 12 | GPT-5.4 (medium) | 2.018 | 1.896 to 2.149 | ||
| 13 | GPT-5.4 (xhigh) | 2.012 | 1.905 to 2.134 | ||
| 14 | Claude Opus 4.7 (high)Stories completed: 347/400 | 1.815 | 1.696 to 1.923 | ||
| 15 | Claude Sonnet 4.6 Thinking 16K | 1.705 | 1.611 to 1.812 | ||
| 16 | Xiaomi MiMo V2.6 Pro (thinking)Stories completed: 375/400 | 1.617 | 1.469 to 1.754 | ||
| 17 | Claude Opus 4.6 Thinking 16K | 1.178 | 0.995 to 1.338 | ||
| 18 | Claude Opus 4.8 (xhigh) | 0.808 | 0.710 to 0.910 | ||
| 19 | Muse Spark 1.1 (high) | 0.804 | 0.714 to 0.910 | ||
| 20 | Muse Spark 1.3 (high) | 0.624 | 0.539 to 0.718 | ||
| 21 | DeepSeek V4 Pro (high) | 0.590 | 0.483 to 0.720 | ||
| 22 | GLM-5.2 (max) | 0.505 | 0.419 to 0.600 | ||
| 23 | Grok 4.7 (high) | 0.444 | 0.236 to 0.641 | ||
| 24 | GPT-5.2 (medium) | 0.383 | 0.214 to 0.546 | ||
| 25 | Claude Opus 4.8 (high)Stories completed: 399/400 | 0.294 | 0.204 to 0.397 | ||
| 26 | Gemini 3.8 Flash (high) | 0.266 | 0.139 to 0.401 | ||
| 27 | Qwen 3.8 MaxStories completed: 398/400 | 0.258 | 0.137 to 0.368 | ||
| 28 | Kimi K2.6 | 0.128 | 0.028 to 0.235 | ||
| 29 | Muse Spark 1.2 (high) | 0.000 | -0.091 to 0.102 | ||
| 30 | MiniMax-M3 | -0.013 | -0.136 to 0.117 | ||
| 31 | Mistral Medium 3.1 | -0.335 | -0.453 to -0.192 | ||
| 32 | DeepSeek V4.1 Flash (high) | -0.407 | -0.529 to -0.273 | ||
| 33 | DeepSeek V4 Pro Preview | -0.490 | -0.591 to -0.384 | ||
| 34 | Qwen3.8-27BStories completed: 389/400 | -0.562 | -0.742 to -0.393 | ||
| 35 | Xiaomi MiMo V2.5 Pro | -0.628 | -0.765 to -0.480 | ||
| 36 | Qwen 3 Max Preview | -0.640 | -0.808 to -0.491 | ||
| 37 | Gemini 3.7 Flash (high) | -0.653 | -0.748 to -0.537 | ||
| 38 | Qwen 3.6 Max Preview | -0.893 | -1.018 to -0.754 | ||
| 39 | GLM-5.1 | -0.996 | -1.184 to -0.831 | ||
| 40 | Kimi K2.5 Thinking | -1.039 | -1.203 to -0.853 | ||
| 41 | Xiaomi MiMo V2 Pro | -1.178 | -1.413 to -0.945 | ||
| 42 | Baidu Ernie 5.1 | -1.198 | -1.348 to -1.058 | ||
| 43 | Mistral Large 4 (high) | -1.585 | -1.698 to -1.458 | ||
| 44 | Mistral Large 3 | -1.720 | -1.836 to -1.603 | ||
| 45 | Gemma 4 31B Reasoning | -1.840 | -1.941 to -1.742 | ||
| 46 | ByteDance Seed2.0 Pro | -1.936 | -2.060 to -1.802 | ||
| 47 | Gemini 3.5 Flash | -1.939 | -2.043 to -1.843 | ||
| 48 | Gemini 3.1 Pro Preview | -2.139 | -2.251 to -2.027 | ||
| 49 | Qwen 3.6 Plus | -2.232 | -2.406 to -2.053 | ||
| 50 | Mistral Medium 3.5 | -2.509 | -2.635 to -2.379 | ||
| 51 | Qwen 3.7 Max | -2.614 | -2.748 to -2.485 | ||
| 52 | DeepSeek V3.2 | -2.829 | -3.075 to -2.565 | ||
| 53 | Grok 4.6 (high) | -2.833 | -2.967 to -2.678 | ||
| 54 | GPT-OSS-120B | -3.130 | -3.249 to -3.006 | ||
| 55 | MiniMax-M2.7 | -3.719 | -3.845 to -3.573 | ||
| 56 | Grok 4.3 | -4.208 | -4.376 to -4.025 | ||
| 57 | Grok 4.5 (high) | -5.040 | -5.135 to -4.956 |
57 of 57 models and settings shown. These rankings apply to this test and publication date.
Go straight to the matchup.
Compare two models on the work they actually did against each other.
This is the share of matched story comparisons won by A, with ties counting as half. It uses only prompts both models answered and can differ from the overall writing ranking.
How the scores work.
The writing score combines judges’ preferences across story comparisons. Higher scores mean the AI judges preferred this model’s stories more often, compared with this group of models. Zero is the average of this group, not a passing grade.
The calculation uses a Thurstone rating. The lines show 95% uncertainty ranges, estimated by repeating the calculation with different samples of stories and judges.
6 models finished fewer stories than planned. Their results show how many stories were completed.
The uncertainty ranges describe the comparisons that were completed. They do not account for missing stories or judgments.
Results as of 2026-10-07. View the original publication ↗