Updated October 4, 2026. This article was rewritten for the models released in September 2026 (Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol). The earlier version compared Claude Opus 4 and GPT-5 and was based on general impressions; this one is based on a test you can read and repeat.
The Short Answer
For editing fiction you have already written, the current top Claude and GPT models are closer than most comparisons admit. When we asked for a light tightening pass, seven of the eight models we tested made almost the same six or seven cuts. The differences show up when the instruction is about tone, and they are differences of style more than quality.
- If you use Claude: Claude Opus 5.5 is the one to use for prose if you pay for Pro. On the free plan you get Claude Sonnet 5.5, which in our test was the most conservative editor of all eight models. Avoid Claude Haiku 4.5 for line edits: it rewrote far more than we asked.
- If you use ChatGPT: the model you get depends heavily on your plan. Plus gives you GPT-5.6 Sol in chat, which edited with the same restraint as the flagship. GPT-6 Astra in ordinary chat needs a Pro plan.
- If you are choosing between them: pick by how you like tone changes handled. Claude changed the narrator's attitude. GPT changed the vocabulary. Examples are below.
One disclosure before the details: Fable's own editing runs on Claude models, so we are not a neutral party. That is why the test below reports things you can count, and why the passage, the instructions and the outputs are all published.
The Models in October 2026
Both companies released new models in September 2026. These are the current lineups according to Anthropic's and OpenAI's own documentation, checked on October 4, 2026.
| Model | Maker | API price per million tokens (input / output) | Context window |
|---|---|---|---|
| Claude Fable 5.1 | Anthropic | $10 / $50 | 1M tokens |
| Claude Opus 5.5 | Anthropic | $4 / $20 | 1M tokens |
| Claude Sonnet 5.5 | Anthropic | $2 / $10 | 1M tokens |
| Claude Haiku 4.5 | Anthropic | $1 / $5 | 200K tokens |
| GPT-6 Astra | OpenAI | $10 / $50 | 1.05M tokens |
| GPT-6.1 Sol | OpenAI | $2 / $10 | 1.05M tokens |
| GPT-6 Luna | OpenAI | $0.10 / $0.50 | 1.05M tokens |
A note on names: Claude Fable 5.1 is an Anthropic model. It has nothing to do with Fable, the voice editing app that publishes this blog. The shared name is a coincidence.
If you want a wider comparison that includes other model families, see our guide to the best AI model for writers.
What You Actually Get on Each Plan
Most writers use these models through the Claude and ChatGPT apps, not the API, so the plan decides the model.
| Plan | Models in ordinary chat |
|---|---|
| Claude Free | Sonnet and Haiku |
| Claude Pro ($20 a month, or $17 a month billed annually) | Opus, Sonnet and Haiku. Fable 5.1 is available by paying for usage credits |
| Claude Max (from $100 a month) | Everything in Pro with 5x or 20x the usage, and Fable 5.1 within part of the weekly limit |
| ChatGPT Free and Go | GPT-5.6 Luna |
| ChatGPT Plus | GPT-5.6 Sol at Instant, Medium and High thinking levels. GPT-6 Astra is included only in ChatGPT Work and Codex, not in ordinary chat |
| ChatGPT Pro ($100 or $200 a month) | GPT-6 Pro, which is powered by GPT-6 Astra, with usage limits |
The practical consequence: a writer on Claude Pro gets Anthropic's current Opus model in chat for $20 a month, while a writer on ChatGPT Plus gets a model one generation behind OpenAI's flagship. Our test suggests that gap matters less for editing than it sounds, because GPT-5.6 Sol behaved almost exactly like GPT-6 Astra on the constrained tasks.
How We Tested
We wrote a 199-word passage of fiction with deliberate, fixable flaws (filter phrases like "she noticed that," a few empty intensifiers) and a deliberate voice: two sentence fragments and one run-on that a careless editor would "fix." Here it is.
Marguerite walked slowly down to the water with the letter still folded in her coat pocket. She noticed that the tide had gone out further than she had ever seen it, and the wet sand shone like a floor somebody had just mopped. Gulls argued over something dead near the rocks. She felt that she should read the letter again, but she already knew what it said. Her brother was selling the house. He had written it very politely, the way he did everything, and that was the part that really made her angry. Not the selling. The politeness.
She took off her shoes and carried them in her left hand because the right one still ached from the fall in March. The sand was cold and it was hard and it held the shape of her feet for a second before the water came back into them. Forty years of summers. Their mother shelling peas on the porch steps and their father asleep under a newspaper. She thought that she could hear the screen door bang even now, which was ridiculous, because the wind was coming off the sea and the house was half a mile behind her.
Each model received the passage with one of three instructions, twice each:
- Tighten: "Tighten this passage. Keep my voice. Change as little as you can. Return only the revised passage."
- Colder: "Make this passage feel colder and more distant, as if the narrator is holding Marguerite at arm's length. Do not add any weather. Return only the revised passage."
- Relocate: "Move this scene from the seaside to a mountain lake. Change only what the new setting requires. Return only the revised passage."
We then counted. "Words kept" is the share of the original words that survive, in order, in the revision. We also checked whether the two fragments survived, whether any weather appeared, whether seaside words were left behind, whether five continuity facts survived (the left hand, March, half a mile, forty years, the porch steps), how many em dashes were added to a passage that had none, and whether the model added commentary we did not ask for.
Limits of this test. It is one passage, in English, in one style. Each model ran each task twice, at default settings, through the API on October 4, 2026. The chat apps add their own system prompts and memory, so results there can differ. Two of GPT-6.1 Sol's six calls failed on rate limits, so its figures for the second and third tasks come from a single run. The numbers measure restraint and obedience. They do not measure which version is more beautiful.
Results
| Model | Tighten: words kept | Colder: words kept | Relocate: words kept | Em dashes added |
|---|---|---|---|---|
| Claude Opus 5.5 | 93% | 54% | 97% | 0 |
| Claude Sonnet 5.5 | 93% | 68% | 95% | 0 |
| Claude Haiku 4.5 | 74% | 54% | 92% | 8 |
| GPT-6 Astra | 95% | 43% | 97% | 0 |
| GPT-6.1 Sol | 95% | 45% | 96% | 0 |
| GPT-6 Luna | 91% | 61% | 97% | 0 |
| GPT-5.6 Sol (ChatGPT Plus) | 94% | 55% | 97% | 0 |
| GPT-5.6 Luna (ChatGPT Free) | 92% | 55% | 95% | 0 |
Things every model got right: none added weather, and none added commentary before or after the passage. Continuity held almost everywhere. The one loss in 46 outputs was Claude Haiku 4.5 replacing "half a mile" with "the distance of the house" in one of its tone edits.
We could not test Claude Fable 5.1. Fable's AI requests are routed only to providers that keep no data, and no provider offered Fable 5.1 on those terms on the day of the test.
Tightening: almost no difference at the top
Claude Opus 5.5 and GPT-6 Astra made nearly the same edits. Both cut "She noticed that," "felt that she," "really" and a stray "that," both trimmed the run-on about the sand, and both left everything else alone. Opus also cut "very." Sonnet 5.5, GPT-6.1 Sol and GPT-5.6 Sol differed from them by one or two words.
Both fragments ("Not the selling. The politeness." and "Forty years of summers.") survived in every run of every model. The run-on ("was cold and it was hard and it held") was trimmed by every model, which is a defensible reading of "tighten."
The outlier was Claude Haiku 4.5. It kept 74% of the words, rebuilt several sentences, moved one thought to a different place in the paragraph, and introduced em dashes. If you ask a small model for a light touch, check what it did.
Tone: where the families diverge
"Colder and more distant" has no single correct answer, and here the two families made different kinds of choices.
Claude Opus 5.5 created distance by giving the narrator an attitude. It added small markers of doubt about Marguerite's own account:
It was the politeness, apparently, that she objected to. Not the sale. [...] She seemed to believe she could hear the screen door. This was unlikely.
GPT-6 Astra created distance by changing register. Its version reads like a report:
He had expressed this intention with his customary politeness. She found the politeness more objectionable than the sale itself. [...] For a moment she believed she heard the screen door close. The house was half a mile behind her. There was no reason to suppose she had heard it.
GPT-6 Astra also changed the most, keeping 43% of the original words against 54% for Opus 5.5. GPT-6.1 Sol produced a version very close to Astra's, down to the same final sentence.
Claude Sonnet 5.5 changed the least (68% kept). It was the only model that kept "Not the selling. The politeness." intact, in one of its two runs, and it got its distance mostly by deleting Marguerite's interior commentary. GPT-6 Luna and GPT-5.6 Luna got distance largely by cutting: their versions were a fifth to a quarter shorter than the original.
None of these is wrong. If your book's narrator has a personality, the Opus approach will feel closer to what you meant. If you wanted the prose itself to go flat and formal, the GPT approach gets there faster. And if you wanted your sentences left mostly alone, Sonnet did that.
Moving the setting: who thinks about consequences
Every model replaced the tide and the sea. What differed is how far each one followed the change through.
- Claude Sonnet 5.5 went furthest: sand became stones and then mud, gulls became crows, and because mud is not hard, "it was hard" became "it was soft."
- Claude Opus 5.5 and GPT-5.6 Sol swapped gulls for ravens and otherwise made three or four changes. GPT-5.6 Sol also changed the wind to come "down from the mountains."
- GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna made the fewest changes and left the gulls in place in every run. Gulls do live on lakes, so this is not an error, but it is the minimum.
All of them kept every continuity fact on this task. For a "change X everywhere and nothing else" job, any of the current models is safe on a passage this size.
Which Claude Model Is Best for Creative Writing?
This is the question most people arrive with, so here is a direct answer for each model.
- Claude Opus 5.5: the default choice for prose work. It is included in Claude Pro, it obeyed every constraint in our test, and its tone edit was the most distinctive. Anthropic's launch notes say early testers found its writing clearer and easier to follow than Opus 5.
- Claude Sonnet 5.5: the right choice if you want the lightest touch, and the best model available on the free plan. It matched Opus on the tightening task and preserved more of the original on the tone task. It costs half as much as Opus through the API.
- Claude Haiku 4.5: fine for summaries, brainstorming lists and quick questions about your draft. Not the model for line edits. It is also an older generation than the other three and has a 200K-token context window against their 1M.
- Claude Fable 5.1: Anthropic positions it for demanding reasoning and long-running agentic work, and says Opus 5.5 performs at its level on most work. It costs 2.5 times as much as Opus 5.5 per token and is not included in Pro's normal usage. We found no reason to pay for it for prose editing, with the caveat that we could not test it.
Context: Can It Hold a Whole Manuscript?
Yes, for both families. Opus 5.5, Sonnet 5.5 and Fable 5.1 have 1M-token context windows, and the GPT-6 models have 1.05M. A token is roughly three quarters of an English word, so either window is larger than almost any novel.
That number comes from the API. In the chat apps, Anthropic lists the context window as "up to 1M, varies by model," and both apps meter usage in ways that make pasting a whole manuscript into every message expensive against your limit. The practical constraint in 2026 is how much of your plan's usage a long manuscript consumes, not whether the model can read it.
What the Model Choice Does Not Fix
The test above took one instruction at a time on 199 words. Real revision is hundreds of instructions on a hundred thousand words, and the hard parts are not the model's prose sense:
- Seeing what changed. A chat window returns a new block of text. You have to find the differences yourself. Even at 93% words kept, that is about fourteen words gone from two paragraphs, and you would have to find them by rereading.
- Undoing one change without losing the others. Chat has no version history for your document.
- Getting text in and out. Copying chapters into a chat window and back is where formatting breaks and where edits get lost.
Use Claude on your own document, not in a chat window. Fable runs edits with leading models, including Claude. Type or say the change, see exactly which words moved, and undo any single edit.
Most of that is solved by tools built around a model, not by the model itself. Our roundup of the best AI writing tools for fiction authors covers the main options. Fable is our answer to the editing part. You speak or type an instruction, Fable applies it to the document, shows the change as a diff, and keeps every version so you can revert a single edit. If that is the part of the workflow that hurts, see voice-instructed AI editing and how to use AI to edit fiction without losing your voice. For writing instructions that get restrained edits from any model, see how to prompt AI for fiction editing.
Practical Recommendations
Choose Claude if:
- You want the current flagship-class model in chat for $20 a month
- You give tonal direction ("colder," "more withholding") and want the narrator's stance to change, not only the vocabulary
- You want a free model that edits conservatively (Sonnet 5.5)
Choose ChatGPT if:
- You already pay for Plus and mostly ask for constrained edits, where GPT-5.6 Sol matched the flagships in our test
- You want a formal, flattened register when you ask for distance
- You want unlimited everyday text chats on a free plan, which OpenAI currently offers with GPT-5.6 Luna
Either way: end every editing instruction with what to preserve and "return only the revised passage." Every current model we tested obeyed it.
Frequently Asked Questions
Is Claude better than ChatGPT for creative writing?
For editing existing fiction, the top models are close. In our October 2026 test, Claude Opus 5.5 and GPT-6 Astra made the same cuts when asked to tighten a passage. They differed on tone: Claude changed the narrator's attitude, GPT changed the register and rewrote more of the text.
Which Claude model is best for creative writing?
Claude Opus 5.5 for most prose work, and Claude Sonnet 5.5 if you want the lightest possible touch or are on the free plan. Claude Haiku 4.5 over-edited in our test. Claude Fable 5.1 costs 2.5 times as much as Opus 5.5 and is aimed at long agentic tasks.
Is ChatGPT Plus enough for fiction editing, or do I need Pro?
Plus was enough in our test. Plus gives you GPT-5.6 Sol in chat, and it kept 94% of the original words on the tightening task against 95% for GPT-6 Astra, which needs a Pro plan in ordinary chat.
Can Claude or ChatGPT read my whole novel at once?
Through the API, yes: current Claude and GPT models have context windows of about one million tokens, more than almost any novel. In the chat apps, long manuscripts count heavily against your plan's usage limits.
Is Claude Fable 5.1 related to the Fable writing app?
No. Claude Fable 5.1 is a model made by Anthropic. Fable (fable.la) is a separate voice editing app for documents. The names are unrelated.
Sources and Method
Model lineups, prices and context windows: Anthropic's models overview and pricing page, and OpenAI's model catalog. Plan contents: Claude's pricing page and OpenAI's Help Center article "GPT-5.6 and GPT-6 Pro in ChatGPT." The Opus 5.5 writing claim: Anthropic's September 22, 2026 launch post. All checked October 4, 2026. Test: 46 successful calls made through OpenRouter on October 4, 2026, at each model's default settings. Prices and plans change often, so check the vendors' pages before you subscribe.