How much does it cost to write a novel with AI?

Every AI-writing pricing page tells you "starts at $20 a month" and refuses to answer the question you actually have, which is how much the whole novel is going to cost you before you finish it. The honest version requires token math. Token math is not difficult, and the answer is more reassuring than the marketing copy suggests. Here is the real number for a 100,000-word novel, with the assumptions stated, so you can run the numbers for your own project.

The output side: how many tokens is a novel

A token is roughly three quarters of a word in English. A 100,000-word novel is therefore around 130,000 to 150,000 tokens of final prose output. Different tokenizers split slightly differently. Claude, GPT, and Gemini tokenizers all land in the same ballpark, within 10% of each other.

So if you generated every word of a 100,000-word novel from an AI, and only once, you would buy 130,000 to 150,000 output tokens. That is the floor.

In practice, you do not pay for the floor. You pay for the full pipeline, which includes input context, multiple drafting passes, edit passes, audits, and the revisions the model produces that you eventually discard.

The full pipeline: input plus output

A complete novel-drafting pipeline like the one inside Pendraic does the following per scene:

• Loads the relevant Story Index entries, World Index entries, the scene plan, the voice profile, and the prior scene's tail into the input context. This is anywhere from 5,000 to 20,000 input tokens per call depending on how rich the project's structured data is.
• Drafts the scene as output. A 2,000-word scene is about 2,700 output tokens.
• Runs a humanize and polish pass. Roughly the same scene length back in as input, similar length back out as output.
• Optionally runs an audit pass to flag continuity issues. Mostly input cost.

Across a full novel of fifty to sixty scenes, with drafting plus one polish pass plus periodic audits, the total token usage lands in the range of 1.5 to 3 million combined input and output tokens. The variance is mostly about how many revision passes you take and how dense your structured data is.

Current model pricing

Pricing as of mid-2026, per million tokens:

• Claude Sonnet 4.6: around $3 input, $15 output.
• GPT-5: around $1.25 input, $10 output.
• Gemini 2.5 Pro: similar to GPT-5, slightly lower on input.

These numbers move. Check the provider pages before you commit. What does not move much is the ratio: output is roughly five to ten times the cost of input on every frontier model.

The worked example

Take a 100,000-word novel drafted through the full pipeline, using Claude Sonnet 4.6 throughout, with a single polish pass and one mid-draft audit. Token totals approximately:

• Input tokens: 1.4 million. Cost at $3 per million: $4.20.
• Output tokens: 600,000 (drafting plus polish plus audit summaries). Cost at $15 per million: $9.00.
• Total: about $13.20.

That is the low end, assuming reasonably tight prompting and minimal regeneration.

A more realistic case for a working novelist who runs multiple drafting passes on the harder scenes, revises freely, and uses both Sonnet for drafting and a cheaper model for cleanup:

• Input tokens: 3 million. Cost at blended $2 per million: $6.00.
• Output tokens: 1.2 million. Cost at blended $12 per million: $14.40.
• Total: about $20.40.

The full range, accounting for projects that run heavy on revision and audit cycles, lands between $25 and $80 from blank page to finished draft. The high end assumes you are regenerating freely, running deep audits, and using the most expensive model for everything.

If you switch to GPT-5 or Gemini 2.5 Pro for drafting, the same novel comes in 30 to 50% cheaper. Output tokens dominate the cost, and those models are cheaper on output.

Compared to flat subscriptions

The other model for AI-assisted novel writing is a flat monthly subscription. Sudowrite, Novelcrafter, and similar tools charge $20 to $30 per month for various tiers. Pendraic's Managed AI path also sits on this side of the line, with credits sold as wallets.

The honest comparison:

• Six months of Sudowrite at $20 per month: $120.
• Six months of Novelcrafter at $30 per month: $180.

Many writers finish a draft inside three months. Many do not. If you draft fast and the subscription tracks the calendar rather than usage, you may pay $60 for a $25 novel. If you draft slowly across a year, you may pay $240 to $360 for what would cost $50 to $80 in pay-as-you-go.

There is also a tier ceiling on most subscriptions. The $20 tier on most tools caps you at a usage limit that a serious revision pass blows through in a week. Upgrading to the unlimited tier is usually $60 to $100 per month. At that price the BYO-AI math wins decisively.

The reverse case: BYO-AI gets pricier than a subscription if you regenerate compulsively. Every time you tell the model "try again," you are spending output tokens. Writers with a clean intent and a willingness to take the first decent draft pay closer to the low end. Writers who chase perfection on every paragraph pay closer to the high end.

Where the cost goes on a real project

It is worth knowing where the money goes, because that is where you have leverage.

• Output tokens dominate. Anything that reduces unnecessary output reduces cost. A tool that drafts a scene once, well, with good structured context, costs a fraction of a tool that drafts the same scene three times because the first two were unusable.
• Audits are cheap. A long-context audit pass reads the manuscript but produces a small structured report. Input is cheap; output is small. Run audits liberally.
• Voice and humanize passes are not free. Every editorial pass is roughly the cost of another partial draft. Worth it. Budget for it.
• Structured data pays for itself. A drafting call with rich Story Index and World Index context produces better output the first time, which means fewer regenerations, which means lower total cost. The input tokens for context are far cheaper than the output tokens for a bad scene you throw away.

BYO-AI vs Managed AI on Pendraic

Pendraic offers both paths. BYO-AI means you connect your own Anthropic, OpenAI, Google, or OpenRouter key. The numbers above apply directly. You see every charge on your provider dashboard.

Managed AI means Pendraic handles the keys and bills credits from a prepaid wallet. The pricing is slightly higher than raw provider cost because the platform absorbs the operational overhead. It is the right choice for writers who do not want to manage API keys or who write infrequently enough that wallet credits last across multiple projects. The comparison is laid out in detail on the BYOK vs Managed AI page.

For a serious project at full novel length, BYO-AI is usually the cheaper option. For a writer doing short-form work or testing the platform, Managed AI is usually less hassle. Pendraic does not push you toward one or the other.

The honest summary

A full novel through an AI-assisted pipeline lands somewhere between $25 and $80 if you pay providers directly. It lands between $60 and $200 if you go through a monthly subscription, depending on how long the draft takes. Neither price is the cost of the platform itself. It is the cost of the model calls that produce and refine your prose.

The expensive part of writing a novel was never the API bill. It was the time. The model spend, on any honest accounting, is the smallest number in the project.

If you want to see the real per-call cost on your own project, sign in, connect a BYO-AI key, and watch the meter. Every call shows up itemised. No surprises.