
Which AI writes the best book chapters?
Claude writes the strongest long fiction on the one public test built like a book. EQ-Bench's Longform Creative Writing benchmark has each model plan a story, then write it as 8 chapters of about 1,000 words, and a judge model scores every chapter. When we read it on 12 October 2026, Claude Opus 5 led at 86.3 out of 100. Claude Fable 5.1 was next at 85.3, and Claude Opus 5.5 was third at 84.6.
Stamina is the bigger split. The newest flagships from Anthropic, OpenAI and xAI finish chapter 8 within 3% of their chapter 1 score. Google's Gemini 3.1 Pro Preview slips 11%. Budget and older models do worse. Claude Haiku 4.5, Anthropic's budget model from October 2025, drops 29%. Meta's Llama 4 Maverick, an open model from April 2025, drops 45%. So a cheap model that writes a fine opening can fall apart by the last act.
| Selected models (EQ-Bench Longform, read 12 October 2026) | Score out of 100 | Chapter 1 | Chapter 8 | Change by chapter 8 | Slop score, lower is better |
|---|---|---|---|---|---|
| Claude Opus 5 | 86.3 | 17.51 | 18.04 | up 3.0% | 5.64 |
| Claude Fable 5.1 | 85.3 | 17.48 | 17.70 | up 1.3% | 7.59 |
| Claude Opus 5.5 | 84.6 | 17.53 | 17.15 | down 2.2% | 8.97 |
| Grok 4.7 | 83.0 | 16.92 | 17.32 | up 2.4% | 4.88 |
| GPT-6 Sol | 82.8 | 16.71 | 16.54 | down 1.0% | 7.92 |
| GPT-6 Astra | 82.8 | 17.07 | 16.72 | down 2.1% | 9.07 |
| Kimi K3 | 79.6 | 17.02 | 16.20 | down 4.8% | 9.67 |
| Claude Sonnet 5 | 78.3 | 16.16 | 16.00 | down 1.0% | 13.53 |
| Gemini 3.8 Flash | 76.8 | 16.48 | 16.06 | down 2.5% | 27.69 |
| DeepSeek V4 Pro | 75.6 | 16.05 | 15.30 | down 4.7% | 21.83 |
| Gemini 3.1 Pro Preview | 68.2 | 15.57 | 13.92 | down 10.6% | 38.12 |
| Claude Haiku 4.5 | 65.0 | 15.37 | 10.86 | down 29.3% | 14.71 |
| Llama 4 Maverick | 31.9 | 8.02 | 4.40 | down 45.1% | 74.65 |
The table picks 13 of the 140 models on the board, and it leaves out older versions such as Claude Fable 5, which also scores 83.0. The slop column counts the words and phrases models overuse, and it doesn't feed the score. Gemini 3.8 Flash uses three to four times as many as Claude Opus 5.5 or GPT-6 Sol.
One caveat favors Claude. The judge is Claude Sonnet 4.6, and EQ-Bench's methodology page says: "We do not control for the judge possibly preferring its own outputs." Human voters see it as a much closer race. On LMArena's creative writing board, dated 8 October 2026, a preliminary Gemini 4 Argon High scores 1519 and Claude Opus 5.5 High 1517, inside each other's margins. The best GPT entry, GPT-5.6 Sol, ranks 20th at 1468. Gemini 4 Argon isn't on EQ-Bench yet, so nobody has tested it over eight chapters.
How much of your book can each AI hold at once?
If you pay for Claude or Gemini, the app can hold a whole manuscript at once. The free plans of ChatGPT and Gemini don't hold even a short novel, which the Nebula Awards define as "40,000 words or more." Each row below turns tokens into words with the vendor's own rule of thumb, from pages read on 12 October 2026.
| App and plan | Context window | About how many words | Fits a 40,000 word novel | Where a series bible goes |
|---|---|---|---|---|
| ChatGPT Free | 27K tokens | 20,000 | No | Projects, 5 files |
| ChatGPT Go or Plus | 54K tokens Instant, 256K Thinking | 40,000 Instant, 192,000 Thinking | Only on Thinking | Projects, 25 files |
| ChatGPT Pro | 128K tokens Instant, 400K Thinking | 96,000 Instant, 300,000 Thinking | Yes | Projects, 40 files |
| Claude paid plans, Claude Opus 5.5, Sonnet 5.5 or Haiku 5.5 | 1M tokens | 555,000 | Yes | Projects, with retrieval past the window |
| Claude paid plans, Claude Opus 4.7 or 4.8 | 500K tokens | 277,000 | Yes | Projects, with retrieval past the window |
| Gemini without a Google AI plan | 32K tokens | 19,000 to 26,000 | No | Gems |
| Google AI Pro or Ultra | 1M tokens | 600,000 to 800,000 | Yes | Gems |
ChatGPT's number depends on the model you pick. OpenAI's pricing page gives Plus 54K tokens on Instant and 256K on Thinking, so switch to Thinking before you paste a draft. OpenAI counts a token as "approximately three-quarters of a word." Anthropic's context window article lists 1M tokens for Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 on paid plans, and 500K for Opus 4.7 and 4.8. Its model docs put 1M tokens at "roughly 555k words." Google's Gemini limits page gives 32K tokens without a plan and 1 million on Google AI Pro.
A series bible belongs in a project, because a project loads it into every chat. OpenAI's Projects article allows 5 files per project on Free, 25 on Go and Plus and 40 on Pro. Anthropic says Claude Projects are open to free accounts, and on paid plans retrieval can "expand capacity by up to 10x" once project knowledge outgrows the window.
The length of a reply is the other limit. Anthropic's model docs let Claude Opus 5.5 write up to 128K tokens in one API reply, about 71,000 words at Anthropic's ratio. OpenAI's GPT-6 Sol page lists the same 128,000 output tokens. Google lists 65,536 for Gemini 3.1 Pro Preview, about 39,000 to 52,000 words, and xAI's model page gives no output cap for Grok 4.7. None of the chat app help pages we read gives a per reply cap. Ask for one chapter a message anyway. It's how EQ-Bench runs its test, and it's easier to revise.
Is Sudowrite or Novelcrafter better than a chatbot for a novel?
Sudowrite is better for a novelist who wants the tool to do the drafting, because it sells drafting credits and runs its own fiction models. Novelcrafter is better for one who wants to organize a book and pick the model, because its core is a story bible and the AI runs on your own key. Novelcrafter's prose comes from the same frontier models the chatbots use. Sudowrite adds its own.
Sudowrite costs $10, $22 or $44 a month as of October 2026, with 225,000 credits on the cheapest plan and 2,000,000 on Max. Its pricing page says it runs its own fiction models, Muse and Ballad, next to Claude and OpenAI models, and it offers a free trial with no card.
Novelcrafter costs $4 to $20 a month, with a 21 day trial and no free plan, per its pricing page. Every tier above the $4 Scribe plan is bring your own key, so your AI bill goes to OpenAI, Anthropic or another provider on top. Its Codex is the story bible, with unlimited entries for characters and places.
There are two more names that show up on most of the lists. NovelAI says its subscriptions "start at $10/month" for text and image generation. ProWritingAid is for the editing pass: its free plan checks 500 words at a time, and Premium is $30 a month or $120 a year with one Chapter Critique a day.
Whizi, which publishes this page, puts Claude Opus and ChatGPT Sol on one Pro plan at $29.99 a month, and side by side model chats let you draft the same scene in both and keep the better one. A Whizi project carries a series bible of up to 10 pinned files, 32,000 characters each and 120,000 across the project. It loses on manuscript size. Each Whizi turn is capped at 40,000 input tokens, roughly 22,000 to 30,000 words, so a whole draft won't fit in one message the way it does in Claude's own app. Pro's 2,500 credits also buy 125 Claude Opus messages a month.
Can AI write a whole book for you?
Yes, AI can draft a whole book, one chapter per message. It won't keep the plot straight unless you hand it the plan every time, and the drop in the chart above is what happens to a budget model that loses the thread. These five steps keep a draft on track in any chatbot:
- Write the premise, the main characters and the ending yourself, on a page or two.
- Load that page, a chapter outline and a sample of your prose into a project as the series bible.
- Ask for one chapter per message, with the beats it must hit and a word count.
- Paste a three sentence summary of the last chapter before you ask for the next one.
- Rewrite every chapter in your own words before it goes in the manuscript.
Paste this as the chapter prompt and fill the brackets:
"You are drafting chapter [N] of [title], a [genre] novel. Follow the series bible and outline in this project exactly. The last chapter ended like this: [three sentence summary]. This chapter must: [beats]. Write about 1,500 words in close third person on [character], matching the style sample. Don't resolve anything the outline saves for later. After the chapter, list every name, date or fact you used that isn't in the bible."
The last line does the work. A model that has to list what it invented catches its own continuity slips, and you decide whether the bible or the chapter changes. For the prompt habits that keep any draft from sounding generic, our guide to using AI for writing has the full checklist.
Can you copyright or sell a book written with AI?
You can sell it on Amazon, but in the US you can't copyright the parts the AI wrote. The U.S. Copyright Office's January 2025 report on copyrightability concluded that "the outputs of generative AI can be protected by copyright only where a human author has determined sufficient expressive elements." It added that prompts alone don't count.
The same report says AI help doesn't sink a book. Using AI to assist, or putting AI text inside a larger human work, "does not bar copyrightability." When you register, the Office's 2023 guidance says AI content that is "more than de minimis should be explicitly excluded from the application."
Amazon's rule turns on who wrote the words. KDP's content guidelines require you to disclose AI-generated text, images or translations when you publish or republish. Text an AI drafted counts as AI-generated "even if you applied substantial edits afterwards." If you wrote the text and used AI to edit, error-check or brainstorm, KDP calls it AI-assisted, and you don't have to disclose it.
That's what the two agencies have published, and it isn't legal advice. A publishing contract or a rights dispute needs an intellectual property lawyer.
- Draft on a flagship model, not a budget one, for anything past chapter 3
- Switch ChatGPT to Thinking before pasting a manuscript
- Keep the series bible in a project
- Ask the model to list every fact it invented
- Disclose AI-drafted text to KDP, even after heavy edits
Frequently asked questions
Is ChatGPT or Claude better for writing a novel?
Claude, by a small margin on the public tests. Claude Opus 5.5 scores 84.6 on EQ-Bench Longform against 82.8 for GPT-6 Sol. On LMArena's creative writing board it ranks 2nd, while the best GPT entry is 20th, as of October 2026. ChatGPT Plus on Thinking still holds 256K tokens, enough for a full draft. Is Claude Pro worth it covers what the paid plan gives a writer.
What is the best free AI for writing a book?
Gemini and ChatGPT hold about the same on their free plans: 32K tokens without a Google AI plan and 27K on ChatGPT Free, roughly 20,000 words each as of October 2026. That covers an outline and a few chapters, not a whole manuscript. Novelcrafter has no free plan, and Sudowrite and NovelAI offer trials. Is Gemini free lists what Google's free plan includes.
Do I have to tell Amazon I used AI to write my book?
Yes, if an AI wrote any of the text, images or translations. KDP counts that as AI-generated even after substantial edits. If you wrote it yourself and used AI only to edit, check or brainstorm, KDP calls it AI-assisted and doesn't require you to disclose it. Before you paste a draft into a chatbot, see how to turn off AI training on your chats.
Do cheaper AI models lose the plot in a long book?
Yes, the budget and older ones do. In EQ-Bench's eight chapter test, Claude Haiku lost 29% of its chapter score by chapter 8 and Meta's Llama lost 45%. The newest flagships from Anthropic, OpenAI and xAI stayed within 3%. On a cheaper model, hand it your outline every chapter. Best AI model for every task helps match model to job.