Writing an 80,000-word novel with AI costs between about 9 cents and $47 in raw model fees, depending on two things: which model writes it and how much of the book you re-send with every request. A cheap model fed only what each scene needs finishes the whole draft for under a dollar. A premium model in one long chat can cost fifty times that for the same words.
Those numbers come from list prices and arithmetic you can redo, shown below. Nothing here is a benchmark of quality; a cheap draft you hate costs more than an expensive one you keep.
What 80,000 words costs on seven models (checked November 2026)
Prices are per million tokens, read from OpenRouter's public model list on 11 November 2026. OpenRouter's list shows what each model's provider charges.
| Model | Input $/M | Output $/M | One pass, lean context | Same, 3 tries per scene | One long chat |
|---|---|---|---|---|---|
| DeepSeek V4 Flash | 0.09 | 0.18 | $0.09 | $0.27 | $1.03 |
| Claude Haiku 5.5 | 0.10 | 0.50 | $0.13 | $0.40 | $1.17 |
| DeepSeek V4 Pro | 0.88 | 1.76 | $0.89 | $2.68 | $10.06 |
| Grok 4.7 | 2.00 | 6.00 | $2.24 | $6.72 | $23.07 |
| Claude Sonnet 5.5 | 2.00 | 10.00 | $2.67 | $8.00 | $23.49 |
| GPT-5.6 Terra | 2.00 | 12.00 | $2.88 | $8.64 | $23.71 |
| Claude Opus 5.5 | 4.00 | 20.00 | $5.33 | $16.00 | $46.99 |
Source: model prices from OpenRouter (11 Nov 2026), costs computed by KudoWrite from the assumptions below. Fees for adding credits and any provider surcharge are not included.
The assumptions, so you can change them
- Length: 80,000 words, about 106,700 tokens at the common rule of thumb of 4 tokens to 3 words. Your language and style will move this a little.
- Scenes: 200 requests of about 400 words (530 tokens out) each.
- Lean context: every request sends about 4,000 tokens of input: a short story-so-far, the cast and facts the scene needs, the previous passage and your instruction.
- Three tries: you keep one of three generations per scene, so all costs triple. Most writers regenerate; many don't triple.
- One long chat: you keep a single conversation going. Each request resends everything before it, so the input grows from 3,000 tokens to over 100,000 tokens by the end: 11.2 million input tokens in all, against 0.8 million for lean context.
The formula is output tokens × output price + requests × context tokens × input price. Put your own numbers in.
Why the long chat costs ten times more
Output is the same in every column: the novel. The difference is input. In a single chat the model rereads chapter 1 when it writes chapter 30, and you pay for the reread each time. On Claude Sonnet 5.5 that turns $2.67 into $23.49 for identical words. Providers cache repeated input at a discount on some models, but the cache must stay warm, and you cannot count on it across a writing session with breaks.
The second cost of the long chat is quality, not money: past a point the model loses track of details buried in the middle, which is the problem behind the AI forgetting your story. Sending a short summary plus only the bible entries a scene touches is cheaper and usually more accurate.
What the subscriptions charge instead (checked November 2026)
| Plan | Price | What it is |
|---|---|---|
| Sudowrite Hobby & Student | $10/month | 225,000 credits a month |
| Sudowrite Professional | $22/month | credits sized for a heavier month (see its page) |
| Novelcrafter Hobbyist | $8/month | the app; its AI features use your own provider key, which you pay separately |
| Claude Pro | $20/month billed monthly | a general chat with usage limits |
A subscription's price is fixed and your usage is not. Writing 80,000 words in a focused month, a $10–22 plan can be a bargain; the same plan costs the same in the three months you write nothing. Sudowrite says its credits don't map directly to words, so we can't price a novel in them and won't pretend to.
Cost per beat: the number to track
Across the table, cost per 400-word scene runs from a twentieth of a cent (DeepSeek V4 Flash, one try) to under three cents (Opus 5.5, one try). Raw model fees are rarely what stops a draft. What does: a context so thin the characters change eye colour, or so fat you pay ten times. Watch cost per scene on your first chapter, then multiply by 200.
Where KudoWrite sits
KudoWrite is the lean-context column made into an app. It keeps the story so far and a story bible that update themselves as you write, picks the entries each passage needs, and sends only those, so you don't buy the reread. There is no subscription: you buy credits in one-off packs when you write, $3.50 for 100, $7.50 for 500 (about a full book) or $12.50 for 1,000 (a well-revised one). A credit balance doesn't reset monthly, so a month off costs nothing. An Expand is about a fifth of a credit; a full beat, with context upkeep, about one.
Those packs include everything KudoWrite does around the model: the bible, summaries, the edit pass and export to EPUB and PDF. Raw model fees in the table above are the floor; the pack is what you pay to have the book kept consistent for you.
How to price your own book in ten minutes
- Write one scene of 400 words in whatever tool you're testing.
- Note the input and output tokens the tool reports, or the credits it spent.
- Multiply by your scene count (word target ÷ 400), then by how many tries you usually need.
- Compare with the monthly price of any subscription you'd otherwise use.
If you want to see a real figure for your own genre, start a book in KudoWrite and watch the balance move as the first scene lands.