A chapter costs the same as a sentence.

Long-form is where per-character pricing bites hardest, because length is the entire product. A book is a few hundred thousand characters by definition, so a meter prices the work by the thing you cannot reduce.

2,000

characters a request — a book is a loop over them

The arithmetic of length

A single request carries up to 2,000 characters and streams gap-free, so a chapter is a loop and a book is a longer loop. On an unmetered seat the loop costs the same, whatever its length — which means a 400-page book and a 40-page one have the same synthesis bill.

The practical effect is on revisions. Narration gets corrected: a mispronounced name, a paragraph the author rewrote, a chapter that reads too fast. Re-rendering those is free here, so corrections are a normal part of the process rather than a budget conversation.

Eleven hours, one voice

Film

Long-form is where a character meter is felt hardest, because the product is length.

The controls a long read needs

temperature sets melodic range — measured 10.1 to 18.0 semitones across its span, which is the difference between a recited chapter and a performed one.
cfg_weight loosens the pacing, so a reflective passage gets room instead of being hurried.
A pronunciation dictionary fixes proper nouns once per request — character names, places, invented words — rather than after a listener complains.
Break tags place the beat between sections exactly where the manuscript wants it.

Notes — an engineer's checklist

01How do I keep a voice consistent across a whole book?

The voice is a request parameter, so every chapter is conditioned on the same reference and there is no drift between sessions. Send a seed as well if you need a specific take to be exactly reproducible when you re-render a corrected paragraph.

02What happens to the bill on a very long book?

Nothing. Characters and minutes are unmetered on a seat, so length changes how long the render takes and not what it costs.

03Can I fix a single mispronounced word without re-rendering the chapter?

Yes — send that passage again with a pronunciation_dict entry and splice it in. With a fixed seed the surrounding audio is unchanged, so the join is clean.