Beginner's Guide to Choosing a Writing Model: September 2026
You can choose the latest version for models of the same series, or you can always use the version you like.
Select by task
| What are you going to do | Best choice | Value for money |
|---|---|---|
| Unpacking the book | Unpacking the chapter: Qwen 3.8Max / glm 5.3 Analysis and interpretation: Gemini 3.1 Pro | Unpacking: Qwen 3.7plus Analysis and Interpretation: DeepSeek 4.0 Pro |
| World view, structural design | Opus 5 | Qwen 3.8Max / glm 5.3 |
| Creativity, brainstorming | Gemini 3.8 Flash Gemini 3.1 Pro | DeepSeek V4 flash DeepSeek V4 pro |
| Write text | Gemini 3.8 Flash Gemini 3.1 Pro Kimi K3 | DeepSeek V4 flash DeepSeek V4 pro |
| Review and modification suggestions | GPT 5.6 Sol | glm 5.3 |
Three important suggestions
- **Put down the outline first, and then write the main text. ** In the Plan stage, the outline of the first chapter is saved as a file. In the Act stage, you can only write according to it; if you want to make changes, change the chapter outline first, and you are not allowed to realize the subsequent plot in advance across chapters.
- **Straight text cannot be delivered without editing. ** No model in the evaluation can directly score more than 80 points; the evaluation of the best model is only "it can be released after a little refinement".
- **Summary the context before changing the model, and then open a new window. ** Make good use of
SOLOENT.md, otherwise the new model will read too much context at once and the effect will be significantly worse.
The text needs attention
**Fixed sentence pattern. ** "Not... but" "not only... also" "suddenly" "suddenly" and "subconsciously" were independently named by multiple reviewers. The dashes are used more than ten times in the model chapter - this is the first sign for readers to identify AI.
**Open the eyes of heaven. ** After the model read the outline, he asked the character to tell information that he should not know, and the logic of the character's thinking and behavior was completely wrong. This is the structural problem that hurts the reading experience the most.
**Beginning and ending. ** The middle paragraph is usually fine. The description of the environment at the beginning and the sublimation summary at the end are the most likely to expose the AI flavor.
**Chapter boundaries. ** 8 of the 18 models in the evaluation wrote the plots of subsequent chapters into the first chapter in advance. The rhythm of the opening chapter was out of control, and there was no content to write in the second chapter.
The above conclusions come from our extensive and ongoing model evaluation. The evaluation project is completely open source: benchmark.soloent.ai