As of November 6, 2026, the best AI for writing is a tie between Claude Opus 5.5 and Gemini 4 Argon. They hold the top two places in Arena's Creative Writing votes, 1,520 and 1,517, and their margins overlap.
ChatGPT's best-rated model on the same board, GPT-6.1 Sol, sits 66 points lower. If you want one cheap model for volume work, Gemini 3.8 Flash scores level with Claude Fable 5.1 at a fraction of the price.
Those votes come from people comparing two anonymous answers to the same prompt and picking the one they prefer. It is the closest public signal to "which one writes better", and it still measures taste on short prompts, not whether a model keeps your voice across a 3,000-word draft. So the picks below are where to start, and the last section shows how to check them on your own work.
Claude and Gemini lead the blind writing votes
We read Arena's text leaderboard on every data refresh. The table below is the November 6, 2026 snapshot for models you can use today: Creative Writing is the vote on stories, essays, and other open-ended prose, and Instruction Following is the vote on prompts with explicit rules. The ± figure is Arena's confidence margin, and two rows whose ranges overlap are a tie.
| Model | Creative Writing | Instruction Following | API price, input/output per 1M tokens |
|---|---|---|---|
| Claude Opus 5.5 | 1,520 ±13 | 1,522 ±10 | $4/$20 |
| Gemini 4 Argon | 1,517 ±13 | 1,525 ±10 | $2/$10 |
| Claude Fable 5.1 | 1,489 ±11 | 1,500 ±9 | $10/$50 |
| Gemini 3.8 Flash | 1,487 ±7 | 1,491 ±6 | $0.75/$3.75 |
| Muse Wonder (preview) | 1,487 ±15 | 1,498 ±12 | $4/$20 |
| Qwen3.8 Max (open weight) | 1,469 ±9 | 1,475 ±7 | Varies by host |
| Claude Sonnet 5.5 | 1,468 ±14 | 1,486 ±11 | $2/$10 |
| GPT-6.1 Sol | 1,454 ±14 | 1,486 ±12 | $2/$10 |
We left two rows out on purpose. Gemini 4 Carbon scores 1,515 ±20 but Google is releasing it to a limited group of cyber defenders first, and GPT-6 Astra (1,450) is rolling out first to enterprises in OpenAI's Trusted Access Program. Neither is something most writers can open today. Prices are from our pricing registry, read November 12, 2026.
Three things in this table change the usual advice. Claude no longer leads alone: Gemini 4 Argon matched it within weeks of its September 30 release. Price barely tracks the vote at the top, since Gemini 3.8 Flash ties Fable 5.1 while costing far less per token. And the newest OpenAI models are strong at following rules (GPT-6.1 Sol ties Sonnet 5.5 on Instruction Following) but trail on open-ended prose.
ChatGPT or Claude for writing? Claude wins most blind votes
Compared app to app, using the best model each one offers, Claude and Gemini lead and ChatGPT trails on prose. The table and sentence below update from the live board on every rebuild:
| Arena category | ChatGPT | Claude | Gemini |
|---|---|---|---|
| Creative Writing | GPT-6 Astra, 1,450 ±11 | Claude Opus 5.5, 1,518 ±11 | Gemini 4 Argon, 1,525 ±10 |
| Instruction Following | GPT-6.1 Sol, 1,483 ±10 | Claude Opus 5.5, 1,519 ±9 | Gemini 4 Argon, 1,529 ±8 |
Gemini 4 Argon (1,525 ±10) and Claude Opus 5.5 (1,518 ±11) are tied at the top of Arena's Creative Writing votes, because their margins overlap. ChatGPT's GPT-6 Astra (1,450 ±11) is further back. Arena scores as of November 9, 2026.
What a gap in points means is easy to misread. On Arena's rating scale, a 66-point lead predicts the higher model wins about 59% of head-to-head votes, not every one. That is an expectation from the ratings, not a count of these two models' direct matchups: roughly two times in five, the model expects voters to prefer the ChatGPT answer. So the vote says Claude is the better default for prose, and it also says a single sample on your own prompt can go either way.
Two caveats change how to use this. The rows are API models, and an app plan can route you to a smaller or faster model unless you pick one, so check which model your plan actually runs. And the gap narrows on rule-bound writing: on Instruction Following, GPT-6.1 Sol trails the leaders by about 39 points. If your work is forms, templates, and fixed-length variants, ChatGPT is closer than the prose vote suggests. For the full app comparison beyond writing, including plans and limits, see ChatGPT vs Claude.
Which AI to use for each kind of writing
The votes cover short, open-ended prompts. Most real writing jobs add something the vote doesn't test: a source pack, a template, a brand voice, or a story bible. Here is where we would start for each, and the case where each pick loses.
Blog posts and long-form articles
Start with Claude Opus 5.5 or Gemini 4 Argon, whichever app you already use. Both lead the prose vote and both accept a 1M-token context, enough for an outline, a source pack, and your style guide in one prompt. Run the other one beside it on the same brief once; the tie means your brief decides.
The pick loses when the article is mostly synthesis from PDFs, URLs, and data files. Then source handling matters more than prose, so test how each model reads your files and links before you judge its sentences.
Editing and rewriting
Claude Opus 5.5 is our first trial for edits that must keep the author's voice. The test is narrow: give it the original, the style rules, and a list of sentences or facts it may not change, then count every unnecessary rewrite.
Do not ask which edit sounds most polished in isolation. Compare each edit with the source voice and count altered claims, deleted qualifiers, and new clichés. Claude Sonnet 5.5 is the cheaper alternative for routine copyedits, and it ties GPT-6.1 Sol on Instruction Following, so either works when the edit can be written as hard rules.
Copywriting and marketing
Start with GPT-6.1 Sol or Gemini 4 Argon when the batch has fixed fields such as audience, offer, proof, and call to action. Both are close to the top on Instruction Following, which is the vote that rewards keeping to explicit rules. Following the rules does not make copy persuasive, though, and nobody publishes a conversion benchmark.
Hold the offer and evidence constant across models. Reject any variant that invents proof, drops a required qualifier, or changes the offer. Then put Claude Opus 5.5 in the review round when brand voice matters more than machine-readable fields.
Email newsletters and high-volume drafts
Gemini 3.8 Flash is our pick for volume. It ties Claude Fable 5.1 on the prose vote at about a thirteenth of Fable's input price, and its ±7 margin is one of the tightest on the board because it has more than 10,000 votes behind it.
Count edits per accepted message rather than assuming the cheapest generator is the cheapest workflow. A cheap draft that needs a rewrite is not cheap.
Fiction and creative writing
Begin with Claude Opus 5.5 and Gemini 4 Argon on the same scene brief and story bible. The Creative Writing vote is the closest thing to a fiction benchmark, but it scores single replies, so the first pass should test continuity: names, ages, locations, injuries, objects, point of view, and what each character knows at that point.
Long context makes more material available to a model. It does not prove the model uses it correctly. A candidate loses when it produces attractive pages that quietly move the timeline or flatten a character's voice. If the story can't leave your machine, Qwen3.8 Max is the strongest open-weight model on the vote.
Developers: docs, READMEs, and technical writing
Start with Gemini 4 Argon when the source pack includes PDFs, URLs, diagrams, and files. Start with GPT-6.1 Sol when the document must follow a fixed schema or template. Test both when the document will guide a production change.
Gate every command, version, default, and compatibility statement against the supplied sources. Search and file access help a model find evidence. They do not certify the sentence it writes.
What the writing scores can't tell you
The Creative Writing vote is the best public signal for prose preference, and it has real limits. Voters judge one reply at a time, mostly on short prompts, without your source pack or your voice. Margins are wide for new models: Gemini 4 Carbon's ±20 comes from 990 votes, while Gemini 3.8 Flash's ±7 comes from 10,581. And a vote rewards the answer people liked, which is not always the answer that was accurate.
Our writing category adds a second signal: instruction-following benchmarks that check whether a model obeys length, format, and exclusion rules. A draft that breaks the brief has failed, whatever it sounds like. But no benchmark we track measures the finished article, the line edit, the campaign, or the chapter. Broad knowledge scores on the full leaderboard have the same boundary: they don't show that a model will make fewer factual errors on your source pack. Check the claims themselves.
Run the same brief before you switch
We have not run this writing comparison ourselves. The following is a reader-run protocol for your own work, not a report of our results.
- Choose three real briefs from the work you publish: one routine, one difficult, and one with a failure you already know well.
- Freeze the brief, source pack, style rules, exclusions, and target length. Change none of them between candidates.
- Remove model names from the outputs and randomize their order before review.
- Apply the factual and constraint gates first. A draft fails if it invents a claim, contradicts the source pack, omits a required point, or breaks a hard instruction.
- Review only the survivors for task quality, then record the minutes and changes required to reach an accepted draft.
| Record | What to check |
|---|---|
| Factual gate | Every checkable claim is supported by the supplied source pack. |
| Constraint gate | Required sections, length, exclusions, format, and calls to action all pass. |
| Drafting | The argument moves, sections do distinct work, and the conclusion follows from the evidence. |
| Editing | The requested changes landed without moving protected facts, qualifiers, or voice. |
| Marketing | Variants differ in a useful way while the offer and proof stay fixed. |
| Fiction | Character state, timeline, point of view, and story-world facts remain consistent. |
| Synthesis | Each important statement can be traced to a supplied source. |
| Editing time | Minutes from raw output to an accepted draft, with the reason for each material edit. |
If cost matters, log the exact API model or consumer plan separately. API tool support does not prove consumer-app access, and token prices do not include the time spent repairing a failed draft.
Step 3 is the same blind comparison Arena runs, applied to your own briefs. If your result disagrees with the public vote, trust yours: it measured the job you actually do. Use the model shortlist to add candidates outside these three companies.
→ Open the model comparison tool · See the writing category
Frequently asked questions
01What is the best AI for writing in 2026?
On November 6, 2026, Claude Opus 5.5 (1,520) and Gemini 4 Argon (1,517) tied for first in Arena's Creative Writing category, where people vote blind between two answers. Their margins overlap, so neither is a clear winner. Start with whichever app you already pay for, then test both on one of your own briefs.
02Is ChatGPT or Claude better for writing?
On Arena's blind Creative Writing votes, Claude is ahead: Claude Opus 5.5 scored 1,520 on November 6, 2026, against 1,454 for GPT-6.1 Sol, ChatGPT's best-rated model there. On Arena's rating scale, that gap predicts Claude's answer wins about 59% of head-to-head votes. ChatGPT is still a sound pick for structured, template-bound copy.
03What is the cheapest good AI for writing?
Gemini 3.8 Flash. On November 6, 2026 it scored 1,487 in Arena's Creative Writing votes, level with Claude Fable 5.1 (1,489), at a fraction of the API price of any model above it. Test it first on high-volume work such as product descriptions, newsletters, or first drafts you will edit anyway.
04Can AI write a full blog post or article?
Current models can produce long drafts, but a large context window does not guarantee continuity, factual accuracy, or publishable voice. Give each candidate the same outline, source pack, style rules, and exclusions. Reject unsupported claims first, then record the editing minutes required to reach an acceptable draft.
05Which open-weight model is best for writing?
Qwen3.8 Max had the highest Creative Writing score of any open-weight model on Arena on November 6, 2026, at 1,469, level with Claude Sonnet 5.5. Its weights are on Hugging Face, so you can run it on your own hardware or through a host when drafts can't leave your infrastructure.