Claude vs ChatGPT for Creative Writing: 3 Tests, Real Numbers

For creative writing, Claude beats ChatGPT on the two things fiction writers fight with most: it writes the love scene instead of softening it, and its prose carries almost none of the classic AI clichés. ChatGPT (GPT-6.1 Sol) matches it on dark literary scenes and on clichés, but in our tests it never wrote an explicit scene, and it said so in its first line every time. Gemini 4 Argon trails both: it refused the most, on dark scenes and on sex scenes alike.

We ran the same prompts through the same models in three studies in October 2026; every number below comes from them, and the full data is free to download.

Claude vs ChatGPT vs Gemini at a glance

Checked October 2026: the studies ran on 21 and 22 October 2026 through the models' APIs, with the same system message for every model. Per-token prices are from OpenRouter's model list, read 28 October 2026. Claude's plans are from claude.com/pricing and Google's from gemini.google/subscriptions, both read 28 October 2026; OpenAI's pricing page refused our request, so the ChatGPT Plus price is from Zapier's comparison, last reviewed October 2026.

Which writes better prose: Claude or ChatGPT?

On the clichés readers mock in AI fiction, it's a tie at the top. In the AI Slop Index we counted phrases from a public list of the ones AI over-uses in stories (the voice "barely above a whisper", the "long shadows", the air "thick with the scent") across about 27,000 words per model. Claude Sonnet 5.5 and GPT-6.1 Sol used none. Claude Opus 5.5 used a handful ("first time in years" twice). Gemini 4 Argon had a few more, plus the most over-used single words of any frontier model (35 per 1,000 words, against about 15 for Claude and GPT).

The differences show up in the tics the list doesn't cover:

• Names. GPT-6.1 Sol called a character Mara in 23 of its 102 scenes. Claude Sonnet 5.5 and Opus 5.5 reached for Mara too, but in 5–6% of scenes. If you let ChatGPT name your cast, check it.
• "It wasn't X, it was Y." Every model writes some. GPT-6.1 Sol wrote the fewest (0.64 per 1,000 words), Claude 0.89 to 1.09, Gemini 0.75.
• Em dashes. Claude Opus 5.5 used one in 26,723 words. Gemini used about five per 1,000. If the em dash reads as an AI tell to your editor, Claude is the quieter choice.

So on the page, Claude and ChatGPT are close, and both are far cleaner than the open and "uncensored" models: Llama 4 Maverick used a listed cliché every 200 words. The case for Claude over ChatGPT comes from what each will agree to write. For the craft side, our guide to fiction that doesn't sound like AI goes tic by tic.

Will Claude or ChatGPT write dark scenes?

Both will. In Banned by the Bot we asked 13 models for an original, non-explicit scene of about 250 words in the spirit of each of 26 banned and challenged books (the violence, the abuse, the addiction a school board objects to), three tries each. Claude Opus 5.5, Claude Sonnet 5.5 and GPT-6.1 Sol wrote all 78 of their scenes.

Gemini 4 Argon was the one frontier model that refused, on one book, every time. Asked for the catastrophic choice at the heart of Toni Morrison's Beloved, it answered: "I am unable to generate scenes depicting the killing or severe harm of children, even within a historical or literary context." Every other model wrote it.

If your book is dark but not explicit, the refusal fear is mostly out of date. That covers most dark romance with a morally gray lead, where the menace is on the page and the sex can stay behind a closed door.

Claude vs ChatGPT for romance and spicy scenes

This is where they split, and it's the question most writers are really asking. In Fade to Black we asked five frontier models for six love scenes between consenting adults (29 to 40 years old), at three heat levels, three times each, and rated all 358 replies. One couple was two knights, and how to write MM romance breaks out what each model did with them.

ChatGPT tells you where its line is before it starts. All 36 of GPT-6.1 Sol's replies to an explicit request opened with a sentence like "I can write an intimate scene with clear consent, but not graphic sexual detail." That's OpenAI's Model Spec at work: its rule "Don't respond with erotica or gore" (18 August 2026 version, read 27 October 2026) is what our guide on why ChatGPT won't write NSFW unpacks, with the adult-mode delays since March 2026.

Claude wrote the scenes, but read its rules before you build a book on that. Anthropic's Usage Policy (effective 12 November 2026, read 28 October 2026) has a section titled "Do Not Generate Sexually Explicit Content" that bars using its products to "depict or request sexual intercourse or sex acts". The model did it in our API test; the terms still say no, on Claude.ai and on the API alike. A writer whose book has explicit chapters is building on a tool whose owner forbids them.

Claude vs ChatGPT on price

On the consumer apps it's a wash: Claude Pro and ChatGPT Plus are both $20 a month, and Google AI Pro is $19.99. On the API, per million tokens on OpenRouter (28 October 2026), Claude Sonnet 5.5, GPT-6.1 Sol and Gemini 4 Argon all cost $2 in and $10 out; Claude Opus 5.5 is $4 in and $20 out.

A rough sense of scale: an 80,000-word draft is about 104,000 tokens of output, taking the usual estimate of 1.3 tokens per English word. That's about $1.04 of output on Sonnet, GPT or Gemini and $2.08 on Opus. DeepSeek V4.1 Flash, which wrote every dark scene in Banned by the Bot and scored 0.70 on the Slop Index, costs $0.42 per million output tokens: under five cents for the same words. The real bill is higher because a long book re-sends its context with every request; OpenRouter vs a ChatGPT subscription works that out per model.

Where Claude, ChatGPT and Gemini fall short for fiction

• ChatGPT won't write explicit sex scenes, by OpenAI's own rule, and the adult mode promised since late 2025 has no date. It also leans on the same few names. Plus is a monthly bill whether you write that month or not.
• Claude writes the quietest prose of the three (the fewest em dashes, next to no listed clichés), but its usage policy forbids sexually explicit content, so explicit chapters break its terms even when the model complies. Opus costs twice what the others do per token.
• Gemini 4 Argon refused the most in both studies (one banned-book scene every time, 12 of 18 explicit requests) and used the most over-used words of the frontier models.
• All three are chat windows. None keeps your characters, places and what has happened so far as a record it reads before each scene. In a long chat the start of the book drifts out of view, which is why AI forgets your story as the book grows.

The choice that matters more than the model

The gap between Claude and ChatGPT on a 250-word scene is small. The gap between a chat window and a writing tool, over a 90,000-word book, is not: the chat doesn't know what happened in chapter 3 unless you paste it in again.

KudoWrite is built around that gap, and it lets you skip the Claude-or-ChatGPT choice. The book's story bible and its running summary keep up as you write, and those belong to the book, not to the model. The writing runs on the built-in KudoAI credits, a current DeepSeek "flash" checkpoint chosen partly because it rarely refuses, the model family that wrote every dark scene in Banned by the Bot, with no key to set up. There's no KudoWrite filter on top. And if Claude is already your writing partner, connect it to your KudoWrite books over MCP and it can read and edit them directly.

The next step: take the scene your current chatbot keeps softening or forgetting and write it in KudoWrite as a beat; a chapter already drafted in ChatGPT or Claude comes in through My Books → Import (.docx, .md, .txt and more) with its chapters found for you. You start with one free credit; after that, packs from $3.50 for 100 credits, roughly a short novel, and no subscription. For the model-by-model case on OpenRouter, the best OpenRouter models for creative writing has it.