What Does It Cost to Write a Novel with AI? We Measured It

Writing a novel with AI costs between nothing and a few dollars a month in model fees, depending on whether the AI writes your prose or helps you write it. Across 42 logged AI generations in NovelCanon, the average paid action cost $0.0078 — and 32 of the 42 cost nothing at all, because they ran on free models.

Search for what AI novel writing costs and you will find numbers that disagree by a factor of ten. One widely-cited estimate puts bring-your-own-key API fees at $20–50 a month for an active drafter, rising to $80–100 for heavy users. Another prices a whole novel at $10 to $180. A third lands at $34–64 a month all-in.

Every one of those pages is published by a company that sells a writing tool, including this one. So rather than add another estimate, we pulled the actual usage log.

Key takeaways

• Cost estimates vary because they price different workflows. An estimate that assumes the AI writes each chapter is not comparable to one that assumes the AI helps you write it. The gap between those two is where almost all the disagreement lives.
• Measured, assistant-style AI actions are cheap. Our logged average was $0.0078 per paid generation, and the single most expensive action we have ever recorded cost 1.3 cents.
• Free models are real, and unreliable. 32 of our 42 generations cost $0.00. They also get pulled or repriced by upstream providers without warning.
• The token math is checkable. We show it below with published prices, so you can run it for your own model and workflow.
• Bring-your-own-key means you pay the provider directly at cost. Compare the models in NovelCanon vs Sudowrite, or start on the free plan.

What actually drives the cost of AI writing?

Three numbers, multiplied together: the price of your model, the number of tokens each request uses, and how many requests you make.

A token is roughly three-quarters of a word. Providers charge separately for input tokens (everything you send — your prose, your characters, the instruction) and output tokens (what the model writes back), and output is typically four to five times more expensive than input. As of publication, Anthropic charges $3 per million input tokens and $15 per million output tokens for its mid-tier model; OpenAI's pricing is structured the same way.

Those are the list prices. With bring-your-own-key, they are also the prices you pay — the software takes no cut of them. That is the whole mechanism: no credits, no conversion rate, no markup between you and the meter.

The interesting question is the middle term. How many tokens does one act of AI-assisted writing actually consume?

What our usage data actually shows

Here is every AI generation logged in NovelCanon between 20 May and 10 August 2026 — 42 requests, across a mix of frontier and free models.

Metric	Measured value
Total generations logged	42
Generations that cost $0.00 (free models)	32
Paid generations	10
Total cost, all 42 generations	$0.078
Average cost per paid generation	$0.0078
Most expensive single generation	$0.0131
Average tokens per generation	665 in / 466 out

Two caveats, stated plainly, because they matter. This is our own usage while building and demonstrating the product — not a novelist drafting a complete manuscript, and 42 requests is a small sample. And roughly three-quarters of those requests ran on free models, which pulls the total down.

But the paid subset is the part that answers the question, and it is unambiguous. Ten paid requests to GPT-4o, Gemini 2.5 Pro, and GPT-5.2 cost 7.8 cents in total. The most expensive single AI action in the entire log — a continuation on a frontier model, sending 2,304 tokens of context and returning 645 tokens of prose — cost 1.3 cents.

At that measured average, a thousand AI actions costs about $7.80. Maxing out NovelCanon's free-plan cap of 20 actions a day, every day for a month, on a paid frontier model, comes to roughly $4.70.

Why do the published estimates land ten times higher?

Because they are pricing a different job, and mostly not saying so.

The most detailed estimate we found assumes each request sends about 3,000 tokens of context and returns about 5,000 tokens of output for a 4,000-word chapter. Read that second number again: 5,000 output tokens is roughly 3,750 words. That assumption is not "AI helped me with this chapter." It is "AI wrote this chapter."

That is a legitimate workflow, and for it the estimate is internally consistent. But it produces a per-request cost roughly ten times our measured average, because it asks the model to produce ten times as much prose. Once you multiply by thirty chapters and a regeneration multiplier, you arrive at a monthly figure that looks alarming — and that has almost nothing to do with the cost of using AI as an assistant.

We can put a name to the gap: the ghostwriter tax. It is not a fee anyone charges you. It is the difference between paying a model to generate your book and paying it to help you write your book, and in token terms it is most of your bill.

The arithmetic is worth running yourself. At $3 per million input tokens and $15 per million output, that 3,000-in / 5,000-out chapter costs about 8.4 cents — nine tenths of it output. Thirty chapters is roughly $2.50 for a complete first-pass draft. Even tripling that for regeneration, a whole AI-generated novel on a mid-tier frontier model lands near $7.50 in model fees.

Which raises an honest point that cuts against the scary numbers and against our own: published per-chapter estimates in the $0.25–$0.40 range run several times higher than what the same token assumptions produce at published prices. We are not claiming anyone is wrong — model versions, retries, and system overhead all move the number. We are saying the arithmetic is short enough to check, and you should check it before accepting any figure, ours included.

So is bring your own key actually cheaper?

Sometimes. The honest answer depends on how you write.

How you work	BYOK model fees	Where a bundled plan wins
A few AI nudges a week	Cents per month, or $0 on free models	Nothing — BYOK is plainly cheaper
Daily drafting with assistance	Low single-digit dollars per month	Nothing meaningful
AI-generated chapters, occasional passes	A few dollars per full draft	Simplicity — one bill, no key
Heavy generation, constant regeneration, premium models	Genuinely climbs; estimate your volume	Predictable ceiling, no bill shock

If you are a heavy generator on frontier models, a credit plan gives you a fixed monthly number and no provider account to manage, and that is worth real money to some writers. Sudowrite's managed model is genuinely simpler, and we said so in our full comparison.

What bring-your-own-key buys is not a guaranteed discount. It is a visible meter. Credits are an abstraction over cost, with a conversion rate the vendor sets and can change; different models burn them at different rates, and it is easy to lose track. With your own key, the number you see on your provider dashboard is the number you paid, and the software's fee is a separate, flat line item.

What about writing a novel entirely on free models?

It works, and we would be overselling it if we stopped there.

OpenRouter maintains a rotating set of zero-cost model endpoints, and they are production models rather than crippled demos. Thirty-two of our forty-two generations used them and cost nothing. On NovelCanon's free plan you can connect a key, point it at a free model, and write indefinitely at no cost at all beyond the 20-actions-per-day cap.

The caveat is stability. The free roster changes constantly — reporting in early August 2026 noted the count dropping from fifteen to fourteen zero-cost endpoints inside a week, with several previously-free model families going offline entirely. Free endpoints also carry rate limits, and any of them can be pulled, repriced, or throttled by its upstream provider without notice.

So: free models are a real way to write a novel with AI for nothing, and a poor thing to build a deadline around. Use them for drafting volume, keep a paid key configured as a fallback, and understand that quality varies more between free models than between frontier ones.

How to estimate your own bill in five minutes

1. Pick your model and find its published price. You want dollars per million input tokens and per million output tokens, listed separately.
2. Estimate tokens per action. A rough rule: words ÷ 0.75. A 200-word continuation is about 270 output tokens; the context sent with it is usually one to three thousand.
3. Multiply. (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price).
4. Multiply by your actual daily volume. Not your imagined volume — count a real writing session.
5. Add the software fee separately. With BYOK these never mix, which is the point.

If the result feels too small, you have probably done it right. The number that surprises writers is not how much AI assistance costs; it is how little, once you stop pricing a ghostwriter you never hired.

The part cost comparisons keep leaving out

Model fees are the cheapest input in this entire equation, and the least important thing to optimise.

What actually costs you is a contradiction that reaches a reader — the character whose eyes change colour in chapter thirty, the subplot you set up and never paid off, the timeline that quietly stops adding up. That is the expensive failure, and no pricing model protects you from it. It is why NovelCanon's Consistency Engine exists and why we treat it, not the token meter, as the reason to be here.

Cost transparency matters because credit anxiety changes how people write — writers ration their revisions when every action feels metered. Remove that and you write the way you should: freely, revising as much as the book needs.

If you want the full picture of using AI without handing over your voice, start with our guide to writing a novel with AI, or compare the field in our honest roundup of AI novel writing software. Or just start free, connect a key, and watch your own meter for a week — it is the only estimate that will ever be accurate for you.