I’ve not used it yet, but SW has the vocabulary feature.
I also like the idea of modes, including custom modes.
And, if I pick AI mode (Super), I can select the AI model. I was not able to capture all the options but this shows some of them.
This is funny, I particularly like the last sentence. 
My recommendation: Claude Sonnet 4.5.
Here is the reasoning.
Your Book Draft mode instructions ask the AI to do something precise and restrained: transcribe faithfully, preserve your vocabulary, handle a small set of formatting commands, and resist the temptation to rewrite. That is an instruction-following task more than a raw intelligence task. It requires a model that reads carefully, follows directions exactly, and does not take creative liberties with your words.
Claude Sonnet 4.5 is the strongest choice for this because it is recognized across the industry for instruction adherence and consistent, predictable output. It will not embellish your prose or substitute synonyms. It handles long-form text cleanly and produces reliable results across repeated sessions, which matters when you are dictating chapter after chapter of a book manuscript. It is also fast enough that the processing delay after you stop speaking will be modest.
GPT-4o would be a reasonable alternative. It is a capable model and would handle the task competently. However, in comparative testing, GPT models tend to be slightly more inclined toward “helpful” rewording, which is precisely what you do not want in this application. For a mode whose entire purpose is to leave your words alone, Claude Sonnet is the more disciplined choice.
Avoid smaller or local models for this mode. Groq offers very fast inference, but the underlying models are less reliable at following complex multi-part instructions. Local offline models (Nano, Fast, Pro, Ultra) are excellent for the standard Voice to Text mode where no AI post-processing occurs, but they lack the sophistication to interpret formatting commands consistently.
In short: select Claude Sonnet 4.5 as the language model for your Book Draft Custom Mode. It will follow your instructions faithfully, process your speech quickly, and keep its hands off your words.