Cost and models
Choosing a model
What to use, and the caching trap that quietly costs real money.
Any of OpenRouter's 400+ models will write you a chapter. The differences that matter are prose quality, price, and — much less obviously — whether the model reuses the chapters it has already been sent.
Sensible starting points
- DeepSeek V4 Flash — cheap, fast, caches properly. Start here.
- GLM 5.2 — stronger prose and continuity for a few times the price, still caches properly.
- Grok 4.5 — arguably the best prose available, with the caveat below.
The caching trap
Each chapter is written with the previous ones in the prompt. Most providers let us reuse that history at roughly a tenth of the price, so chapter 50 costs barely more than chapter 5.
xAI's models are the exception. They only reuse a prompt that is byte-for-byte identical to a previous one. Because every new chapter adds to the history, no two requests ever match, and nothing is ever reused. We measured this on a real 18-chapter novel: about 8% of tokens reused, against roughly 80% on a comparable model. The prose may well be worth it — but it is worth it knowingly.
The cost estimator models this exactly, so you can see the difference in dollars before you commit a novel to a model.
Switching mid-novel
Perfectly fine, and the model picker in the writer toolbar does it per novel. The first chapter after a switch pays a cold cache, then it settles. Voice may shift slightly — the style prompt keeps that much smaller than you would expect.
Free models
Models ending in :free cost nothing but are heavily rate-limited and generally weaker at long-form prose. Fine for experimenting with a premise; frustrating for a real novel.