Claude vs ChatGPT for writing: Claude leads 2 of 3 writing rankings, and GPT-6 Sol beats Opus 5.5 on short stories

Is Claude or ChatGPT better for writing?

Claude is better for writing on two of the three public writing leaderboards, and its lead is widest where people do the voting. Arena's creative writing ranking, published on 15 October 2026, puts Claude Opus 5.5 first out of 414 models. It puts GPT-6 Sol, the model that runs GPT-6 in ChatGPT, at 61st. The gap is there in all four of the writing categories we checked.

Arena shows a person two answers to the same prompt from two models it doesn't name, and the person picks the one they like more. We read every row from the Arena leaderboard dataset, which is the file Arena puts out with each release. The ratings in the chart use style control. Arena's own post says that it adds length and markdown to its model, so that a long or heavily formatted answer can't buy votes on its own. There's one caveat. Opus 5.5 is new, so its rating rests on only 1,945 votes, and the range around it is wide enough that Gemini 4 Argon, in second, could be level with it.

The best OpenAI writer on Arena isn't the one in the chat box. GPT-5.6 Sol at extra high effort is 20th for creative writing, and GPT-6.1 Sol is 30th. You can pick GPT-5.6 Sol in chat on Plus, but OpenAI's model help page says that Plus doesn't include Extra High. And GPT-6.1 Sol only runs in Work and Codex. Arena also tested GPT-6 Sol at max, the top effort in OpenAI's API, while Plus stops at High. So a Plus writer gets a lower setting than the one voters ranked 61st.

What do the three writing leaderboards say?

The three writing leaderboards split two to one for Claude. Both EQ-Bench boards use a Claude model as the judge, and the short story board picks ChatGPT anyway. We built the table below from the data each board had published by 18 October 2026. It compares the two models that each $20 plan puts in its chat window.

EQ-Bench Creative Writing v3 runs each of its 32 story prompts three times, and then a model picks the better of two stories, again and again. Its about page says that Claude Sonnet 4.6 has been the judge for the Elo since 1 March 2026. It also says: "We do not control for the judge possibly preferring its own outputs." That makes the OpenAI lead of 75 points hard to wave away. A Claude judge picked the OpenAI model anyway. The same board's rubric score, graded by the older Claude Sonnet 4, has Claude ahead 16.79 to 16.51, so the Elo is the only short story number GPT-6 Sol wins. GPT-6 Astra is at the top of that board, but on Plus you only get Astra in Work and Codex. And the older Claude Opus 5, which Pro still offers in chat, scores 2,132.6, a hair above GPT-6 Sol.

EQ-Bench Longform has each model plan a story and then write it as eight chapters of about 1,000 words each. Claude wins it by 1.8 points. The judge there is also Claude Sonnet 4.6, so this win carries the self preference caveat that the short story result doesn't. Both of them hold up. Their chapter scores barely drop from the first chapter to the eighth, so neither one falls apart late in a book.

Which is better for each writing task?

Claude wins for fiction, for editing, for long documents and for other languages. ChatGPT wins for short marketing work that needs an image. No public leaderboard scores marketing copy or email on its own, so those two rows rest on what each app can do.

Marketing copy is the clearest win for ChatGPT. Its pricing page lists "More complex and accurate image creation" on Plus, so you can get the headline and the picture from one chat. Anthropic's help center is plain about it: "Claude doesn't generate photos or illustrations the way image-generation tools do." That settles it for ads.

Email goes the same way, for a practical reason. OpenAI's release notes for 8 June 2026 say that if you're on Plus and you've connected Gmail or Outlook, ChatGPT can draft an email and send it without you leaving the chat. At launch that works in English in the US, the UK, Canada, Australia and New Zealand. If it's only the words you want, our AI for email writing guide has prompts that work in both.

Editing goes to Claude. An edit is a test of whether the model does what the brief says, and that's what Arena's instruction following ranking measures. It has Claude second and the OpenAI model 63rd. ChatGPT has the nicer screen to edit in, though. OpenAI added full screen writing blocks for essays, reports and blog posts on 8 June 2026, after it took canvas away from GPT-5.5 on 28 May.

How much text can Claude and ChatGPT read and write at once?

Claude can read about three times as much text in one chat as ChatGPT Plus can. Anthropic's context window article lists 1M tokens for each of its current Opus, Sonnet and Haiku models on every paid plan. Its models overview says that 1M tokens is roughly 555,000 words.

ChatGPT Plus has two windows, and both are smaller. Its pricing page lists 256K tokens on the reasoning models, which it puts at about 320 pages of input, and 54K on Instant, about 40 pages. OpenAI's token guide puts a token at about three quarters of a word, so 256K is roughly 192,000 words. A typical novel fits on a ChatGPT reasoning model but not on Instant. Several drafts plus your notes at once only fit in Claude.

How much each one can write back is a tie, at least on paper. Claude Opus 5.5 and GPT-6 Sol can each write up to 128,000 tokens in one reply through the API. Neither app says what the cap is for one reply in a chat. Our context window comparison covers Gemini and the rest.

Which gives you more control over tone and style?

ChatGPT gives you more tone controls as of October 2026, and Claude is moving its style presets into skills. Going by OpenAI's release notes, there are seven base styles left now that it has retired Nerdy (on 17 March 2026): Default, Friendly, Efficient, Professional, Candid, Quirky and Cynical. Since 19 December 2025 you can also turn four dials, for how warm it is, how enthusiastic it is, how many headers and lists it uses, and how many emoji.

Custom instructions sit on top of all that, in Settings, Personalization. Since 7 November 2025, a change you make there applies to all of your chats at once. That includes the old ones.

Claude's styles are going away in their old form. Anthropic's help article on them was titled "Styles are moving to skills" when search engines last saw it. It said that your custom styles would move to skills that start out switched off, and that the Concise, Explanatory and Formal presets would end. When we checked on 18 October 2026, that page gave a 404. Free, Pro and Max users switch skills on in Customize, Skills, per Anthropic's skills article. What's left for your voice is user preferences and project instructions, which are both on the pricing page.

For your own voice, a project is still the best tool either app has. Put three of your own pieces in it and tell it to keep to their rhythm. ChatGPT's presets set a mood. They don't learn from what you've written.

What do the $20 plans limit for writers?

Both plans cost $20 a month, and neither one tells you how many long drafts you'll get. ChatGPT Plus is the only one of the two that says everyday chat is unlimited.

Anthropic's Pro plan article promises only "More usage per session than the Free plan." Its usage limits article adds that what you do in claude.ai, Claude Code and Claude Desktop all counts toward the same limit. So if you write and code, it's one pool.

OpenAI's Plus help page says: "Currently, we do not support annual billing" for Plus. That makes Claude $40 a year cheaper if you pay up front. Our Claude Pro vs ChatGPT Plus page compares the two plans in local prices.

Where does each one lose?

Claude Opus 5.5 loses the short story test, the one with a Claude judge, and it uses more stock phrases. EQ-Bench gives it a slop score of 10.43 on short stories, against 8.97 for the OpenAI model, and 8.97 against 7.92 on long ones. A lower score means fewer of the phrases that its counter flags as typical of models. Claude also can't make the image for a post. And its style presets are on their way out.

ChatGPT loses wherever people vote. Its chat model is 61st to 72nd in Arena's four writing categories, while Claude is first or second in all of them. Plus also reads about a third of the text Claude can, and it can't pick the efforts Arena tested: Extra High for GPT-5.6 Sol and max for GPT-6 Sol.

Other languages are closer than Arena makes them look. We ran our own Multilingual LLM Bench on 2 October 2026, with 150 questions in 30 languages. GPT-6 Sol got 63.3% of them right and Claude Opus 5.5 got 60.7%. Each of those has a margin of about 7.7 points, so it's a tie. Claude did better on idioms, 45% to 36.7%. In Korean, Arena has no row for Opus 5.5 yet. All three of its top Korean models are from Claude, though, and the OpenAI model is 59th.

Should you use Claude or ChatGPT for writing?

Use Claude for long or careful writing. Use ChatGPT for short copy that has to go out with an image.

The leaderboards don't agree on short fiction, so test that one yourself. Run one of your own prompts in both, and keep the draft you'd have to edit less. Our free AI model comparison tool sends one prompt to several models at once.

Whizi fits one case: you want Claude Opus 5.5 and GPT-6 Sol to write the same draft side by side before you pay for two plans. Side by side chats come with Whizi Pro at $29.99 a month. An Opus message costs 20 credits there and a Sol message costs 10, so the 2,500 credits on Pro cover 125 Opus or 250 Sol messages. One draft from both costs 30. For the ad picture, you'll still want ChatGPT.