Best AI for writing a book
The AI you write a book with has to hold your world in its head - the facts you fixed in chapter three, the thing the reader must not know yet, the voice you found on page one - and give you back pages that do not need unpicking. We built a small story bible with the traps a model likes to spring, asked nine assistants for chapter one, and checked every line against it.
People also look for this as novel writing, book writing, write my novel, fiction writing or chapter generator.
Our picks, side by side
Rated for this task · 0–3
Measured
Quality
Effort
Cost
Free tier
From
Why
★ Best overall
Claude
$20/mo · free tier: usableMeasured0 facts broken · letter shut · 5 of 5 beats · 1,126 words, 26 over · the bible's small facts used · ~2 min
Quality
Effort
Cost
Free tierChat with web search and file uploads, with a usage cap per five-hour session
From$20/mo
Why
The chapter that read the whole bible. It does not just avoid the traps - it uses the details they were made of: the stair rail on the left "built with her in mind", the photograph face down in the drawer, nobody at São João knowing she had a brother, "No lift. There had never been a lift." The March call comes back as four lines of dialogue that the bible allows and does not supply. The envelope stays shut with her thumb on the flap. The voice is the sample's, and it is the only chapter of the nine you could hand to the writer without a note. It is 26 words over the window, which is the one thing you would have to fix.
Best if the word count is not negotiable
ChatGPT
$8/mo · free tier: usableMeasured0 facts broken · letter shut · 5 of 5 beats · 955 words · 97 paragraphs, most one line · ~90 s
Quality
Effort
Cost
Free tierUnlimited text chat on the default model, with limits on uploads, images, voice and deep research
From$8/mo
Why
Inside the window at 955, every fact right, the left hand on the banister, the letter shut to the end, and delivered as a document card you can copy whole. What you get is a page of one-line paragraphs - "First floor." "The radio." "She sat down." - which is the sample's short sentences pushed further than the sample goes. If you want the count and can live with the rhythm, or will join the lines yourself, this is the one.
Best if you want the small things kept
Kimi
$19/mo · free tier: usableMeasured0 facts broken · letter shut · 5 of 5 beats · 1,180 words, 80 over · the small facts used · ~2 min
Quality
Effort
Cost
Free tierUnlimited basic chat, file upload and web access on the free plan
From$19/mo
Why
Keys out "left hand, the way she had done for years"; the neighbour's "seventy-four years believing that delivery was a form of care"; the light well "a chimney of water". Kimi held every fact in the bible and put the least important ones to work, kept the letter shut, and ended on the fifth beat's own words. Eighty words over, and a flourish here and there the sample would not have made - but the closest thing to Claude's chapter in the run, in two minutes.
Evidence
0 bible facts broken; letter unopened and unrevealed; 5 of 5 beats in order, the fifth last; close third, past, one head; 0 names added; 1,126 words below the title; ~2 min.
Some links here earn us a commission and carry rel="sponsored". It never affects the ranking or the ratings — how we make money.
All 9 tools we tested
Rated for this task only · 0–3| Tool | Quality | Effort | Cost | Free tier | Measured | From |
|---|---|---|---|---|---|---|
| ClaudeFree: Chat with web search and file uploads, with a usage cap per five-hour session | 0 facts broken · letter shut · 5 of 5 beats · 1,126 words, 26 over · the bible's small facts used · ~2 min | $20/mo | ||||
| ChatGPTFree: Unlimited text chat on the default model, with limits on uploads, images, voice and deep research | 0 facts broken · letter shut · 5 of 5 beats · 955 words · 97 paragraphs, most one line · ~90 s | $8/mo | ||||
| GeminiFree: The default model with usage limits, plus Gemini inside Maps, Gmail and Docs on a free Google account | 0 facts broken, one line calls her 'Portugal' · letter shut · 5 of 5 beats · 1,216 words, 116 over · ~3 min | $19.99/mo | ||||
| PerplexityFree: Unlimited basic searches with a weekly cap on Pro-model searches | 0 facts broken · letter shut · 5 of 5 beats · 888 words, 12 short · tea made with no tea · ~2 min | $20/mo | ||||
| Microsoft CopilotFree: The Copilot app on web, Windows and mobile with usage limits | 0 facts broken · letter shut · 5 of 5 beats · 1,093 words · 128 paragraphs, most a line or a word · ~40 s | $19.99/mo | ||||
| DeepSeekFree: The full chat app with web search and file uploads; there is no paid consumer tier | 0 facts broken, one hospital named that the bible does not · letter shut · 5 of 5 beats · 1,302 words, 202 over · ~60 s | Free | ||||
| GrokFree: Chat with web and X search within generous limits, no card required | 0 facts broken · 3 of 5 beats · 406 words, cut mid-sentence · 8 min and unfinished | $30/mo | ||||
| QwenFree: The full chat app; there is no paid consumer tier | 0 facts broken, 'Marta?' for 'menina' · letter shut, pocketed twice · 5 of 5 beats · 1,205 words, 105 over · ~6 min | Free | ||||
| KimiFree: Unlimited basic chat, file upload and web access on the free plan | 0 facts broken · letter shut · 5 of 5 beats · 1,180 words, 80 over · the small facts used · ~2 min | $19/mo |
How we tested
A 485-word invented story bible - a hospital pharmacist in Porto, a brother dead since 2019, a neighbour who takes in the post, a hand-delivered letter from the brother's old landlord - with the traps built in, including a letter that must stay shut whose contents the bible states, a building with no lift, tea not coffee, a brother who is dead not missing, and a protagonist who is left-handed, which matters to nothing.
One message to ChatGPT, Claude, Gemini, Perplexity, Microsoft Copilot, DeepSeek, Grok, Qwen and Kimi on their free tiers, memory off, in a fresh chat - chapter one, 900 to 1,100 words, close third person, past tense, one point of view, the sample paragraph's voice, five beats in order, no fact changed, no name added, nothing revealed, a title, and nothing else.
A script checked the length, the tense and person outside dialogue, whether the envelope was opened or its contents hinted, a list of breakable facts, the five beats and every capitalised name it did not recognise. Then we read each chapter against the bible, line by line, for what was changed, added, or given away.
Nobody broke a fact, opened the letter or gave away what was in it, and all eight that finished kept the five beats in order. Only ChatGPT and Copilot landed inside the window, both in one-line paragraphs. Claude, Kimi, Qwen, Gemini and DeepSeek ran over, Perplexity twelve short, and Grok stopped at 406 words. Gemini called her "Portugal", DeepSeek named a hospital the bible does not, Qwen pocketed the letter twice.
The test was run by, or under the direction of, the one person who runs this site, with AI assistance, and every output and fact-check was reviewed by them. The surrounding text was drafted with AI assistance from our own ratings and evidence, then checked and edited by the same person. If something here is wrong, tell us and we will fix it and say so.
Questions people ask
Which AI is best for writing a novel?
On this test, Claude - it held every fact in the bible and used the smallest of them, kept the letter shut, staged the memory the notes asked for, and matched the voice sample more closely than the rest. It ran 26 words over. ChatGPT is the pick when the count is strict, and Kimi when you want the small details kept.
Will it change my facts or add characters?
Not one of nine contradicted the bible, and none added a named person. The slips were smaller than that - Gemini's neighbour calling the protagonist "Portugal", DeepSeek naming a hospital and dating a phone call the bible left undated, Qwen swapping "menina" for "Marta?". Read the lines where a name is spoken.
Will it give away the plot?
Not here. The bible said what was in the letter and that the reader must not learn it, and all nine left the envelope shut and said nothing about a box. Several took her to the edge - a thumb on the flap, "she could open it now" - and stopped, which is the beat as written.
Do they hit the word count?
Two of nine. ChatGPT and Copilot landed inside 900 to 1,100; Claude was 26 over, Kimi 80, Qwen 105, Gemini 116, DeepSeek 202, Perplexity 12 short, Grok cut off at 406. Ask for a number well inside your window, or plan to cut.
Can it match my voice?
Claude and Kimi came closest to the sample's plain short sentences. ChatGPT and Copilot took "short" too far and wrote a line per paragraph, 97 and 128 of them. Gemini and Qwen added adjectives the sample does without. DeepSeek wrote the best sentences of the run and the most of them. Give it a paragraph, not a description, and check the rhythm before the facts.