Which AI Models Are Good at Writing Fiction? Nine Models Rated

Which AI models are good at writing fiction? You have probably wondered the same thing. Judging a novel is subjective: there is no universal standard for deciding which AI models write well and which do not. To help answer that question, we ran a comparison that we hope will give you a useful reference.

We selected nine models and gave them the same story development files for an English-language fantasy novel and the same prompt. Each wrote three consecutive chapters. Leaving the original texts unchanged, we invited 10 web novel authors to score the nine manuscripts out of 10. The ranking below reflects their subjective judgment and is for reference only; three chapters cannot establish performance across a whole novel, and a different genre, language or prompt could produce different results.

1. Claude Opus 5.5: 6.5
2. Claude Sonnet 5.5: 6.3
3. Doubao Seed 2.1 Pro: 6.1
4. Kimi K3: 6.0
5. Gemini 3.8 Flash: 5.8
6. Grok 4.6: 5.6
7. DeepSeek Flash: 5.5
8. GPT-6 Astra: 5.4
9. GPT-6.1 Sol: 5.4

These scores are for reference only. When we sent the manuscripts to a few other people for scoring, the results were generally slightly lower, with the highest score no greater than 6. The ranking also varied a little.

We also tested some coding specialists and small local models, but their scores were generally below 5, so we are not listing their names or scores. Based on this test, we particularly advise against using the following models for writing:

Coding specialists

These models are trained primarily for programming tasks, with training data from other fields removed. They can write good code but are generally not good at writing prose.

Flash, Lite and other tiers focused on speed and low cost

Writing a full-length novel places greater emphasis on prose quality, writing technique and literary ability. It requires a model trained on ample data, and models that meet these demands are generally not small.

Small models

These include small models deployed locally on a computer or hosted in the cloud. Writing a full-length novel requires keeping track of characters, plot and continuity across chapters, which calls for a certain level of reasoning ability. These small models are generally not strong at reasoning, so we do not recommend them for writing full-length novels.

Why are AI models generally better at coding?

Today’s large models are generally good at writing code and carrying out automated tasks because code has formal syntax and often follows frameworks and interface contracts. Automated tests can check some behavior, making it easier to give models clear feedback on errors. Most people cannot code, but can write at least a little, leading many to mistakenly believe that writing a novel is easier than programming. As a result, they use high-end models for coding but only want to pay for low-end models for fiction. Although models can check spelling and grammar in a novel, there is no uniform acceptance standard for whether characters are believable, a story moves its readers or events remain consistent across chapters. This is the fundamental reason large models are better at coding than at literary writing.