The best AI for writing fiction is the one that can see your book

The question arrives in every writing forum, phrased with the confidence of a question that has an answer. Which AI is best for fiction, Claude or ChatGPT or Gemini or Grok? A ranking follows, and the ranking is obsolete within the quarter, and the question is asked again.

We are in an unusual position to answer it, since Proselon works with all of them and sells none of them. You sign in with whichever account you already hold, so we have no engine to defend. Which is how we came to notice that the question is the wrong size. The engine is a real variable — and a small one. The variable that governs the outcome is what the model can see of your book when it writes.

The same engine, two very different collaborators

Consider a model asked to draft chapter thirty.

In the first arrangement it has your last paste, which is to say the preceding scene, a synopsis you wrote in February and have since departed from, and whatever character notes survived the conversation. It writes competently into a void. The prose is fluent, the protagonist is a stranger, the promise you made in chapter nine goes unredeemed because nothing in front of the model records that you made it. You will spend the afternoon repairing this, and much of the repair will consist of telling the model things you have told it before.

In the second arrangement the same model, on the same subscription, reads your outline, the preceding scene, the chapter containing the promise, the pages of your story bible that concern this character, and a thousand words of your own prose at its best. It then drafts. The output is not miraculous, and you will still rewrite it, but you are rewriting a draft of your novel rather than correcting a well-written stranger.

The distance between those two drafts is larger than the distance between any two frontier engines handed the first arrangement. This is the part the rankings cannot capture, because every comparison is conducted in a chat window where every engine is equally blind.

The mechanism is not mysterious. A model knows only what is on its desk at that moment. Voice fidelity likewise depends on exemplars being present at the moment of drafting rather than described in advance, which is the clearest result we have measured. Continuity across a long book depends on the small facts being retrievable rather than remembered. None of these three is a property of the model. All three are properties of the workspace you put it in.

What each engine is good at

Having said all that, the differences are real, and here they are as of July 2026.

Caveat: models update constantly, and every gap below may narrow or invert with the next release.

Claude has been the fiction writers' preference for prose and voice for a couple of years running. Its default register sits nearer to literary prose than to press release, it holds a specified voice longer before drifting, and its dialogue tends to carry subtext instead of announcing feeling. It needs a paid plan, and the free tier will not sustain a drafting session.

ChatGPT is the engine most writers arrive holding, and it earns the position. Instruction-following is excellent, plot logic is strong, and a free account signs in and works within usage limits, which no other frontier engine offers. Its weakness in fiction is the trademark polish, competent and agreeable and uniform, which is why a voice reference matters more here than anywhere else.

Gemini brings capacity and price. It runs on Google's free tier and holds more text at once than its rivals, which matters less than you would suppose, though it does assist a whole-manuscript pass. Of the frontier engines it is likeliest to hand you something needing a rewrite for rhythm.

Grok is capable and the least filtered of the signed-in engines, behind X Premium+ or SuperGrok, which makes it the practical choice for work other providers decline to touch.

Local models through Ollama or LM Studio are free, private and unlimited, at a real cost in prose quality, and they are as much a hardware question as a model one. We treat them properly in Can you write a novel with AI for free?

The conclusion the rankings keep missing

Any ranking is a snapshot, and the leader changes underneath you. The useful inference is not to identify the current winner but to decline to marry your workflow to any engine at all. A tool built on one bundled model binds your book to that model's fortunes. A workspace in which engines are interchangeable lets you follow quality wherever it moves next — and it will move.

And "best" depends on the work in hand. Claude for prose, ChatGPT for structure and instruction-following, Gemini for volume on a budget, Grok for unfiltered material, a local model when privacy is the constraint. A good many writers keep two.

What none of them will do is compensate for a workspace that shows the model two paragraphs of a hundred-thousand-word novel and asks for chapter thirty.

Where Proselon stands

Proselon is engine-neutral by design. The co-writer signs in to your existing Claude, ChatGPT, Grok or Gemini account, or runs a free local model, and each of them receives the identical workspace: your manuscript as plain files on your own computer, chapters it may open and read on demand, a continuity ledger, and a voice reference assembled from your writing. Change engines whenever you please and the book does not move. It is $6 a month, free for 30 days, with every feature and every engine included.

Pick the engine you like. Then give it something to read.