You asked for 5,000 words and received 1,400. You asked again, and it apologized and wrote 1,200 more. Is there a limit?
There are two, and they have nothing to do with each other. One is how much ChatGPT can hold. The other is how much it will write in one reply. For a novelist the second one is the daily annoyance and the first one is the real danger.
We make BookNova, an AI novel writer. Limits are from OpenAI's pages, read on November 22, 2026. Measurements are from our own test on that day and from a published study. Words are estimated at about three quarters of a word per token.
Short answer
- What it can hold: 27K tokens on Free (about 20,000 words), 54K on Go and Plus with the everyday model, 256K with reasoning models.
- What it will write: there is no published word limit per reply in the app. In our own test current models wrote 75 to 110 percent of a 5,000-word request, and Gemini wrote one and a half to two times as much as asked.
- Older models were worse: asked for 8,000 to 10,000 words, GPT-5 with reasoning averaged 9,050 words and GPT-4o averaged 1,241.
- Best method: scenes of 800 to 1,200 words, three or four per chapter.
Limit 1: what ChatGPT can hold
This is the context window, and OpenAI publishes it.
| Plan | Everyday model | In words | Longest single input | Reasoning models | In words |
|---|---|---|---|---|---|
| Free | 27K | about 20,000 | about 12 pages | Varies | Varies |
| Go | 54K | about 40,000 | about 40 pages | 256K | about 190,000 |
| Plus | 54K | about 40,000 | about 40 pages | 256K | about 190,000 |
| Pro | 128K | about 96,000 | about 250 pages | 400K | about 300,000 |
The "longest single input" column explains why pasting your manuscript into a free chat fails at once. OpenAI lists about 12 pages for the everyday model on Free, and about 320 pages for reasoning models on Plus.
The window includes the replies. When it is full you may see "Context window full," and the story's early details begin to go. See why ChatGPT forgets your characters.
Limit 2: what ChatGPT will write
OpenAI does not publish a words-per-reply limit for the app. For developers, the GPT-6 Sol model page lists up to 128,000 output tokens, which would be close to 96,000 words. You will not get that in a chat. Two things are in the way.
Training. Chat models are tuned to be brief and helpful. A request for 5,000 words competes with everything the model learned about good answers, which tend to be short.
The app. Very long replies can be cut or stop at a "continue" button. In September 2026 a Pro subscriber reported on OpenAI's forum that long outputs in copyable blocks were being truncated mid-line.
What we measured: a 5,000-word chapter
We asked four current models for chapter one of a romantic suspense novel, "5,000 words long," three times each.
| Model | Run 1 | Run 2 | Run 3 |
|---|---|---|---|
| GPT-6 Sol | 3,758 | 5,013 | 5,451 |
| GPT-6 Luna (ChatGPT's free-plan model) | 3,783 | 4,935 | 5,131 |
| Claude Sonnet 5.5 | 4,416 | 3,791 | 4,237 |
| Gemini 4 Neon | 10,191 | 9,235 | 7,684 |
Then we asked each for ten chapters of "about 2,500 words" in one conversation, to see if they get shorter:
| Model | Shortest chapter | Longest chapter | Total, 10 chapters |
|---|---|---|---|
| GPT-6 Sol | 2,496 | 3,130 | 27,418 |
| GPT-6 Luna | 2,665 | 4,190 | 32,510 |
| Claude Sonnet 5.5 | 2,275 | 2,981 | 27,012 |
| Gemini 4 Neon | 4,592 | 10,335 | 64,098 |
Two surprises. The chapters did not shrink: chapter ten was as long as chapter one. And the familiar complaint "it writes too little" does not fit these models any more. Gemini has the opposite problem, and one of its chapters even ran into our output cap.
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.
What researchers measured in the previous generation
The ConStory-Bench study asked many models for stories of 8,000 to 10,000 words and counted what came back. These are models of the previous generation, which many complaints online still describe.
| Model | Average length | Share of stories over 8,000 words |
|---|---|---|
| GPT-5, reasoning | 9,050 | 66.4% |
| Claude Sonnet 4.5 | 8,929 | 66.8% |
| Gemini 2.5 Pro | 5,091 | 2.1% |
| Gemini 2.5 Flash | 5,504 | 10.3% |
| GPT-4o | 1,241 | 0% |
(Figures from the paper's length table. Its results table gives Gemini 2.5 Pro a slightly higher average, 5,584.) The lesson is older than these models and still useful: the reasoning model did what it was told, the chat model wrote about a seventh of it. If your chapters come back too short today, check which model is answering before you change the prompt.
Longer is not better
It is tempting to look for the trick that gives you 10,000 words at once. The same study is a good reason not to. Errors grew roughly in line with length, and the contradictions gathered around the middle. A long reply is also expensive to reject: if paragraph four is wrong, the next thirty were built on it.
Most novel chapters run 1,500 to 4,000 words anyway.
How to get a full chapter
The scene method
- Plan the chapter as three or four scenes. For each: where, who wants what, what goes wrong, the last image.
- Ask for one scene at a time, 800 to 1,200 words.
- Accept or fix it before the next. Say "Continue with scene two" and repeat its plan.
- Join them yourself. Write the transitions. Two sentences in your own voice go a long way.
Four scenes of 900 words make a 3,600-word chapter that you have read and approved.
Scene 2 of 4, chapter 7. 1,000 words. Place: the ferry's engine room, night. Ines wants the truth about the loft. Mara wants her to leave. What goes wrong: Tomas walks in. Last image: the toothpick on the floor. Start in the middle of a sentence of dialogue. Do not summarize. Do not end on a thought.
If you need a longer reply
- Choose a reasoning model and say that length is a requirement: "The scene must be at least 1,500 words. Count before answering."
- Give it more to write about. Length follows material. Six beats make a longer scene than two beats and a number.
- Ask for expansion by place, not in general: "Expand the argument on the stairs to 500 words. Leave the rest."
- Ask for a plain continuation: "Continue from the last line without repeating it."
- Avoid copyable blocks for long text. Ask for ordinary text.
Check the count yourself
Models are poor at counting their own words. Paste the reply into your editor and look.
How many replies does a novel take?
| Novel | Scenes of 1,000 words | With one retry in two | Chats, one per chapter |
|---|---|---|---|
| Novella, 30,000 words | 30 | 45 | 10 |
| Romance, 60,000 | 60 | 90 | 20 |
| Thriller, 80,000 | 80 | 120 | 27 |
| Epic fantasy, 120,000 | 120 | 180 | 40 |
At 3,000 words per chapter. It can be done, and it is a project. Can ChatGPT write a whole book? estimates the hours.
What about Claude and Gemini?
Claude's current models hold up to 1M tokens on paid plans. Gemini holds 32K without a plan, 128K on AI Plus and 1M on AI Pro. Reply length varies by model, as the table above shows for their predecessors. Compare them in ChatGPT vs Claude vs Gemini.
If you would rather not count
BookNova has no words-per-reply limit, because it does not work in replies. You set the length of the book and it plans and writes the chapters in order. The limit is your word allowance: one credit is one word, from 50,000 every 30 days. An 80,000-word novel fits in one cycle on the Pro plan. Books finished in the app average about 42,000 words, and about one in ten passes 80,000.
Bottom line: the reply limit is mostly a habit of the model, and scenes get you around it. The window limit is real, and only a better plan, a reasoning model or a fresh chat gets you around that. Stop asking for chapters. Ask for scenes.
FAQ
What is the maximum number of words ChatGPT can write?
There is no published limit per reply in the app. Through the API, GPT-6 Sol allows up to 128,000 output tokens.
How many pages can ChatGPT read?
OpenAI lists about 12 pages per input on Free, 40 on Go and Plus and 250 on Pro for the everyday model. Reasoning models take about 320 pages on Go and Plus and 680 on Pro.
Why does ChatGPT stop in the middle of a chapter?
Either it reached the length it considers complete, or the app cut a long reply. Say "Continue from the last line."
Does Plus write longer answers than Free?
It gives access to reasoning models, which tend to follow length requests better, and a larger window.
Sources, all read November 22, 2026
- ChatGPT pricing and plan limits
- OpenAI: GPT-6 Sol model page
- BookNova test of four models, November 22, 2026 (12 chapters and 40 serial chapters)
- ConStory-Bench, full paper with length tables
- OpenAI community: truncated long outputs, September 2026
- BookNova: count of finished books over 3,000 words, November 22, 2026