All three can write a good page. None of them will write your novel while you sleep. Between those two facts there are real differences: how much of the book each can hold, how it writes, what it refuses, and what you can take out at the end.
We compared the current plans and the public rankings, and left out everything that has nothing to do with fiction.
We make BookNova, which competes with all three for the job of writing a novel. Limits come from OpenAI, Anthropic and Google pages read on November 22, 2026. Rankings are linked and dated, and will change.
Short answer
- Best prose by blind votes: Claude and Gemini are at the top, ChatGPT well behind.
- Most memory for $20: Claude and Gemini, 1M tokens. ChatGPT Plus has 256K with reasoning models and 54K with the everyday one.
- Best free plan for drafting: ChatGPT, for unlimited everyday chats.
- Explicit scenes: none of the three.
- Best overall for a novelist: Claude Pro. Gemini if you want the large window inside Google Docs.
The three side by side
| ChatGPT | Claude | Gemini | |
|---|---|---|---|
| Free plan | Unlimited everyday chats | Limited per 5 hours and week | Limited, 32K window |
| Main paid plan | Plus, $20/mo | Pro, $20/mo | AI Pro, $19.99/mo |
| Cheaper plan | Go, $8/mo | None | AI Plus, $4.99/mo |
| Window on that paid plan | 54K everyday, 256K reasoning | Up to 1M | 1M |
| Keeps files for a book | Projects | Projects | Gems and notebooks |
| Export for writers | Copy, or full data export | Copy, files, full data export | Canvas to Google Docs |
| Explicit sexual content | Restricted | Not allowed | Not allowed |
| Children's picture stories | Text and images by request | Text | Storybook, 10 pages |
US prices. ChatGPT's window sizes are on its pricing page, Gemini's on the limits page, Claude's in its help center.
How much of your novel each one can hold
At about three quarters of a word per token:
| Plan | Window | In words | An 80,000-word novel |
|---|---|---|---|
| Gemini without a plan | 32K | about 24,000 | Less than a third |
| ChatGPT Free, everyday model | 27K | about 20,000 | A quarter |
| ChatGPT Plus, everyday model | 54K | about 40,000 | Half |
| Gemini AI Plus | 128K | about 96,000 | Fits |
| ChatGPT Plus, reasoning model | 256K | about 190,000 | Fits twice |
| Claude Pro, current models | up to 1M | about 750,000 | Fits nine times |
| Gemini AI Pro | 1M | about 750,000 | Fits nine times |
Two things this table does not show. The window includes all your notes and every rejected draft, so real capacity is smaller. And fitting is not remembering. Google itself says that large, detail-heavy prompts can lead to answers that miss connections or details when you exceed the window, and research shows contradictions appear well before that.
Which one writes better?
Blind human votes
Arena shows people two anonymous answers and asks which they prefer. In the creative writing ranking published on November 19, 2026, as read by Whizi:
- Claude Fable 5.2 was first, Gemini 4 Argon second, Claude Opus 5.5 third
- GPT-6 Sol, the model behind GPT-6 in ChatGPT, was 82nd of 424
- The best-placed OpenAI model, GPT-5.6 Sol at its highest effort, was 22nd
Not all of these models are in every app. Fable needs usage credits on Claude Pro, and Google's AI Pro plan is listed with Gemini 4 Neon. Read it as a tendency: Claude and Gemini are close together and ahead.
Benchmarks
EQ-Bench Longform, which asks for eight chapters of about 1,000 words, gives Claude Opus 5.5 84.6 and GPT-6 Sol 82.8. Its short story board puts GPT-6 Sol ahead of Opus 5.5. Both use a Claude model as judge.
Consistency
The only independent study we found on contradictions is ConStory-Bench, from March 2026. It used the previous generation of models, so treat it as history with a lesson:
| Model tested | Errors per 10,000 words | Average length when asked for 8,000 to 10,000 |
|---|---|---|
| GPT-5 (reasoning) | 0.11 | 9,050 |
| Gemini 2.5 Pro | 0.31 | 5,584 |
| Claude Sonnet 4.5 | 0.52 | 8,929 |
| GPT-4o | 0.71 | 1,241 |
Look at the first and last rows. They are both from OpenAI. The reasoning model was the most consistent in the group and delivered the requested length, and the chat model delivered about a seventh of it. Which mode you pick inside an app can matter more than which app you pick.
Our own test of the current models
We gave the same jobs to four current models on November 22, 2026.
| GPT-6 Sol | GPT-6 Luna | Claude Sonnet 5.5 | Gemini 4 Neon | |
|---|---|---|---|---|
| Asked for 5,000 words, got (3 runs) | 3,758 to 5,451 | 3,783 to 5,131 | 3,791 to 4,416 | 7,684 to 10,191 |
| Ten chapters at 2,500 words, average | 2,742 | 3,251 | 2,701 | 6,410 |
| Em dashes per 1,000 words | 1.2 | 1.9 | 0.8 | 8.8 |
| "Like a" similes per 1,000 words | 0.4 | 0.1 | 1.8 | 1.4 |
| Fantasy heroine named Elara, of 20 | 16 | 18 | 19 | 6 |
| Clear breaks of the story rules we checked | 0 | 0 | 0 | 1 |
For the last row we planted facts and rules in a story bible (which hand is scarred, which ear is deaf, when the bridge is closed, no kiss before chapter 7) and searched the ten chapters for them. All four kept nearly everything while the whole conversation was in view. Gemini let the couple kiss in chapter 5.
That is the most useful result here. These models did not forget a 27,000 to 64,000-word story. Forgetting begins when the app stops showing them the beginning, so the size of the window on your plan decides more than the brand.
How we tested: on November 22, 2026 we sent the same prompts to four current models through their APIs with default settings: GPT-6 Sol and GPT-6 Luna (the models behind ChatGPT on paid and free plans), Claude Sonnet 5.5 and Gemini 4 Neon. The apps add their own instructions, windows and limits, so results there can differ. It is a small test, and we make a competing product. We kept all outputs. We searched the chapters for the planted facts and read the passages. A full edit would find more.
What each one refuses
| ChatGPT | Claude | Gemini | |
|---|---|---|---|
| Kissing, desire, tension | Yes | Yes | Usually |
| Closed-door love scenes | Yes | Yes | Usually |
| Explicit sex | Restricted, adult mode paused | No | No |
| Violence that serves the story | Mostly | Mostly | No "gratuitous" violence |
This table is our reading of the three policies. Details and quotations are in Can AI write spicy romance?
Which to use for which job
| Job | Pick | Why |
|---|---|---|
| Brainstorming for hours | ChatGPT Free | No limit on everyday chats |
| Outline and structure | ChatGPT reasoning or Claude | Good at logic and holes |
| Drafting scenes | Claude | Instruction following, subtext |
| Reading the whole manuscript | Claude or Gemini AI Pro | The 1M window |
| Finding contradictions | ChatGPT reasoning or Claude | Ask for quotes as proof |
| Working in Google Docs | Gemini | Canvas exports straight to Docs |
| A short illustrated story for a child | Gemini | Storybook |
| Blurb, ads, emails | ChatGPT | Fast, and makes images too |
A workflow with two chatbots
Use one as writer and another as editor. Draft in Claude, then give the chapter to ChatGPT or Gemini with a cold instruction: "You did not write this. List the five weakest sentences, any contradiction with the attached canon, and one place you stopped believing it." Saying so invites a franker reading.
One warning: do not let the editor rewrite. Take the list and go back to the first chat. Otherwise the book ends up with two voices.
What none of them do
None of the three has a chapter list, a timeline, a record of who knows what, or a way to turn the result into a KDP paperback. You are the system. That is fine for writers who like it.
If you do not, look at writing apps that keep the state for you, or at BookNova, which plans and writes the whole draft in order, tracks characters and threads, allows explicit content where the genre needs it, and exports PDF, EPUB and DOCX depending on plan. It is one payment. Side by side: ChatGPT vs BookNova and Claude vs BookNova.
Verdict: if you will pay for one, make it Claude Pro for the prose and the window. Choose Gemini AI Pro if you live in Google Docs. Keep ChatGPT Free next to either for ideas. If you pay for ChatGPT already, switch to a reasoning model before you conclude that it cannot hold a story.
FAQ
Which is best for writing a book for free?
ChatGPT for unlimited chatting, Claude for style. Expect to restart conversations often on both.
Is Gemini good at creative writing?
Yes. Its Gemini 4 models are near the top of the Arena creative writing ranking. The free plan's 32K window is the weak point. More in Gemini for Writing a Novel.
Which has the longest memory?
Claude's current models on paid plans and Gemini AI Pro, both up to 1M tokens.
Can I move a book from one to another?
Yes, through exported text. Gemini even has an import for chat history from other assistants. See how to move and finish it.
Sources, all read November 22, 2026
- ChatGPT pricing, Claude pricing, Google AI plans
- Gemini Apps limits and Claude context windows
- Whizi: Claude vs ChatGPT for writing, Arena and EQ-Bench data
- BookNova test of four models, November 22, 2026
- ConStory-Bench, full paper
- Go and Plus prices: OpenAI on ChatGPT Go and press coverage