Best LLM for writing: Claude, GPT-6, Gemini, Grok and DeepSeek ranked by job.

1. The ranking: the best LLM for each writing job.

Which AI is best for writing depends on what you are writing. A model that ranks 4th on short fiction can rank 11th on matching an editorial voice, and paying flagship prices for meta descriptions costs up to a hundred times what the cheapest current model charges for the same output. Here is the pick for each job as of October 5, 2026, with the evidence in the sections that follow.

If you only want one model for everything, use Claude Opus 5.5. It is the model Claude's developer docs tell you to start with, it sits near the top of the only cross-vendor writing board, and it costs less than either flagship above it. The rest of this page is the case for splitting the work, and for spending your own time where no model helps.

2. Every current model, and where you can use it (October 2026).

The app you pay for decides which of these models you can actually write with. The table lists the models that matter for writing on October 5, 2026, with its release date, where you get it, its API price per million input and output tokens, its context window and its knowledge cutoff, from each vendor's own docs.

Four things the table cannot hold:

• The chat window is not the API. In ChatGPT, GPT-6.1 Sol, GPT-6 Sol and GPT-6 Luna run in Work and Codex, not in Chat, where paid plans still write with GPT-5.6 Sol and the Free and Go plans with GPT-5.6 Luna. GPT-6 Sol itself (September 22) has been replaced by GPT-6.1 Sol at the same price.
• Plans gate the best models. Claude Opus 5.5 needs a paid Claude plan; Claude Sonnet 5.5 has been the free default since September 28. Gemini 3.8 Flash needs Google AI Pro or Ultra in the app, and from October 9 the free Gemini plan is Flash-Lite only.
• Prices move. The Gemini Flash price doubles to $1.50 and $7.50 on January 1, 2027. GPT-5.6 Sol's $4 and $20 is a promotion through at least November 21. DeepSeek charges half outside its peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays).
• Announced is not available. Claude Haiku 5.5, DeepSeek V4.1 Pro and Gemini 4 Argon (open first to cyber defenders in Google's Fairwind Program) have no public date, and GPT-6.1 Sol has no date for ChatGPT Chat. They are listed below, not ranked.

How the lineup got here:

1. 2026-09-01 · Claude Fable 5.1.
2. 2026-09-02 · Gemini 3.8 Flash.
3. 2026-09-04 · GPT-6 Astra, public.
4. 2026-09-10 · DeepSeek V4.1 Flash; V4 Pro set to be phased out, then kept on the API after users objected.
5. 2026-09-21 · Grok 4.7.
6. 2026-09-22 · Claude Opus 5.5; GPT-6 Sol and GPT-6 Luna.
7. 2026-09-28 · Claude Sonnet 5.5, the new default on the free Claude plan.
8. 2026-09-29 · GPT-6.1 Sol, in ChatGPT Work and Codex.
9. 2026-09-30 · Gemini 4 Argon announced, not open to writers.
10. 2026-10-09 · Free Gemini plan drops to Flash-Lite.
11. 2026-10-14 · GPT-5.5 retires from ChatGPT.
12. NEXT · Claude Haiku 5.5, DeepSeek V4.1 Pro, wider Gemini 4 Argon access: announced, no dates.

This page ranks across vendors. For the full lineup inside each family, including older models still on sale and what each plan includes, see our guides to the best Claude model for writing, which ChatGPT model writes best and choosing a Gemini model for writing.

3. How we compared them.

A ranking is only as useful as the method behind it, so here it is in full, including what we did not do.

• Scope. Five model families: Claude, ChatGPT's GPT models, Gemini, Grok and DeepSeek. A model counts as current if a paying customer could use it on October 5, 2026. Announced models are listed, not ranked.
• Specs from the source. Prices, context windows, knowledge cutoffs and plan access come from each vendor's own documentation: Claude's developer docs, ChatGPT's model docs, Google's Gemini API docs and model cards, xAI's Grok 4.7 docs, and DeepSeek's API pricing page and release notes. Where a vendor page was silent or would not load, launch coverage filled the gap.
• Writing quality from named evaluations. Three public ones, each covering a different job. Lech Mazur's Short-Story Creative Writing Benchmark has 56 models write 600 to 800-word stories that must work ten required elements into one piece; its September 26 update added Opus 5.5, Gemini 3.8 Flash, Grok 4.7 and DeepSeek V4.1 Flash. Louis-François Bouchard's editorial-voice benchmark, as reported by Decrypt on September 6, ranks models on editorial writing in a house voice. Definition's AI writing leaderboard (March 30, 2026) scores short business writing. None covers every model or every job.
• What we did not do. We did not run our own blind panel, so no score on this page is ours. Where no public evaluation covers a job (editing, non-English writing), the pick is a judgment from documented specs, and the section says so.
• Tie-breaks, in order. Writing evidence first, then knowledge cutoff, then price per article, then whether a writer can get the model in an app they already pay for.
• Cost math. List price times tokens, before thinking tokens. Claude's current tokenizer uses about 1.8 tokens per English word (its docs put one million tokens at roughly 555,000 words); for the others we use the 75 words per 100 tokens rule from ChatGPT's help center, which is an approximation for Grok and DeepSeek.

Two models outside the five families place high on the short-story board: GLM-5.3 in 6th and Kimi K3 in 9th. Both are open-weight and worth a test if you already use them; they are not ranked here. And because the evidence is thin on some jobs, the most useful thing this method gives you is a shortlist. Run your own brief through the top two picks for your job before you commit a content program to either.

4. Best LLM for blog writing and long-form articles.

For a 1,500 to 3,000-word article written to a brief, the pick is Claude Opus 5.5, for three reasons a writer can check.

• It sits near the top of the cross-vendor board. On Lech Mazur's short-story benchmark, Opus 5.5 places 3rd of 56, behind Claude Fable 5.1 and the older Claude Opus 5 at its highest effort setting. The best model from any other vendor is GPT-6 Astra in 4th. DeepSeek V4 Pro is 21st, Grok 4.7 23rd, Gemini 3.8 Flash 26th.
• It knows the most recent world. Its reliable knowledge runs to June 2026, the latest of any model in the table.
• Its launch notes single out writing. The Opus 5.5 launch notes describe its writing as clearer and easier to follow, with the most important information up front, and say its documents need less editing before you share them.

Price is the last check. One standard article (a 3,000-word brief in, a 2,000-word draft out) costs about 9 cents on Opus 5.5 through the API. Here is the same article on every current model, at list price before thinking tokens:

Why not Claude Fable 5.1, which tops the board? Its score there is 3.796 against 3.767 for Opus 5.5, it costs two and a half times as much per token, and Claude's docs rate it the slowest model in the lineup. Keep it for long research turned into one finished document, or fiction where the last few points matter. Why not GPT-6 Astra? On fiction it is the best non-Claude model. On matching a house voice, the evidence runs the other way: Bouchard's editorial-voice benchmark ranked Astra 11th, five places below GPT-5.6 Sol, and Decrypt's review called it mediocre at work where the standard is taste. At $10 and $50 per million tokens it is also the most expensive model here, with Fable.

On a budget, Gemini 3.8 Flash drafts the same article for about 1.3 cents, knows the world to March 2026 and is the best-placed Gemini model on the board. Grok 4.7 sits between the two at $2 and $6, with knowledge to May 2026, but xAI pitches it at coding, agents and knowledge work, and its 23rd place gives a writer no reason to pick it over either. If you are choosing between the two biggest assistants rather than models, our Claude vs ChatGPT writing comparison sets out a same-brief test you can run on both.

5. Best model for editing and rewrites.

Editing is a different job from drafting. An edit pass reads a finished draft and returns it tighter, closer to your house voice, or stripped of the habits that make AI prose easy to spot, and a real article takes several passes. Speed and price per pass matter as much as raw quality, and no public benchmark scores editing directly, so this pick is a judgment from documented specs.

The pick is Claude Sonnet 5.5. It shipped on September 28 with the same June 2026 knowledge as Opus 5.5 at half the price ($2 and $10 per million tokens). Its launch notes list producing polished documents among its strengths and say it runs more than 30% faster than Sonnet 5, and it is the default model on the free Claude plan, so anyone can try it. One edit pass on a 2,000-word draft with a page of house rules costs about 4.5 cents through the API.

The runner-up is the model most ChatGPT subscribers already have open. GPT-5.6 Sol, which paid plans get in Chat, placed 6th on the editorial-voice benchmark, five places above GPT-6 Astra, and ChatGPT's August update gave it more direct responses and tighter formatting. Its catch is age: its knowledge stops at February 16, 2026, so keep it on style and away from facts.

On any model, the instruction matters more than the pick. Claude's prompting guide notes that a general request to avoid a generic AI look mostly swaps one default for another, so name the specific patterns instead: the phrases, paragraph shapes and punctuation you never use. Our checklist for making AI drafts sound human lists the patterns worth naming, and the guide to writing with Claude covers keeping those rules in every chat.

6. Best model for outlines and briefs.

A brief decides what an article covers, in what order, for which search query and from which sources. It is reasoning over material rather than prose, so the writing boards say little about it. What counts is how well a model reads a long pile of sources, holds all of it at once and turns it into a plan.

The pick is GPT-6.1 Sol in ChatGPT Work. ChatGPT's model docs describe it as near-Astra performance for complex work at a lower cost; through the API that is $2 and $10 per million tokens, a fifth of GPT-6 Astra's price, with a 1.05 million token context window and knowledge to April 2026. In ChatGPT it runs in Work and Codex from the Plus plan up, not in Chat. On a budget, Gemini 3.8 Flash reads a million tokens of sources for 75 cents through the API, and the Google AI Pro plan that includes it in the app also puts Gemini inside Google Docs.

No model supplies the two inputs a brief for search depends on. The query has to come from measured demand: there is no keyword database inside any model, so any search volume it quotes is a guess. And the angle has to come from you: your customers, your numbers, your first-hand experience. Our brief-first content process shows what goes into a brief before a model sees it, and the guide to blog SEO covers picking queries a small site can actually rank for.

7. Best model for titles and meta descriptions at scale.

Titles and meta descriptions are short, repetitive and numerous: a 400-page site needs 800 of them. Consistency and cost matter more than the last bit of polish per line, so this is the one job where the cheapest capable model wins.

The pick is GPT-6 Luna, which ChatGPT's docs call their most efficient model for focused, high-volume tasks. At $0.10 per million input tokens and $0.50 per million output, with knowledge to May 18, 2026, a title and a description for each of 400 pages (a 1,500-word excerpt of each page in) costs about 9 cents in total through the API. In ChatGPT it runs in Work and Codex; the closest model in Chat, on the Free and Go plans, is GPT-5.6 Luna.

The runners-up are Gemini 3.5 Flash-Lite ($0.30 and $2.50, on every Gemini plan) and DeepSeek V4.1 Flash ($0.30 and $1.20 at peak, half that off-peak). Claude's cheapest model, Haiku 4.5 ($1 and $5), is still on sale, but its knowledge stops at February 2025 and Claude Haiku 5.5 is announced without a date. For short copy where the line itself has to sell (headlines, taglines, product names), test a flagship instead: Definition's March 2026 business-writing leaderboard put Gemini 3.1 Pro first at 7.63 out of 10, though that was before most current models existed.

At scale, the model is rarely what goes wrong. Titles drift past the width Google shows, two pages end up with the same description, and nobody notices for months. A purpose-built meta description generator and an SEO title tag generator check length against what Google displays, which a chat window does not.

8. Best model for writing in other languages.

All three writing evaluations named on this page score English prose. There is no public ranking we would trust for long-form writing in Spanish, German, Japanese or Chinese, so the honest answer is a test, and a cheap one. This pick is a judgment.

What the docs do say: Claude's developer docs list multilingual capability for every current Claude model, which tells you the model is built for it but not how well it writes your language. Two practical points follow. The words-per-token rules above are for English, and many languages need more tokens per word, so budget from a sample in your language rather than the English math. And if you write in Chinese, add DeepSeek V4.1 Flash to the shortlist: it comes from a Chinese lab, it is the cheapest drafter here after GPT-6 Luna, and DeepSeek's chat app is free to test in.

The test takes an afternoon. Shortlist two models: the winner of your English test (for most people Claude Opus 5.5) and a cheaper drafter (Gemini 3.8 Flash). Write the brief in the target language rather than translating an English one, and give both models the same brief plus 500 words of native prose you like. Then have a native-speaking editor mark every line they would change, without knowing which model wrote which draft. Keep the model with fewer marks, and run the test again when either vendor ships a new version.

Watch for the failures a fluent non-native reader misses: the wrong register (tú or usted, du or Sie), English punctuation habits, idioms translated word for word, and examples, prices and brands from the US market dropped into a local page.

9. Knowledge cutoffs: what each model does not know.

A knowledge cutoff is the date a model's training data stops. Past it, the model either says nothing or guesses, and a guess reads exactly like a fact. As of October 5, 2026:

• Claude Opus 5.5, Sonnet 5.5 and Fable 5.1: June 2026, the reliable-knowledge date in Claude's docs.
• Grok 4.7: May 2026. GPT-6 Luna: May 18, 2026.
• GPT-6 Astra and GPT-6.1 Sol: April 30, 2026.
• Gemini 3.8 Flash and 3.5 Flash-Lite: March 2026, with some domains limited to January 2025.
• GPT-5.6 Sol and GPT-5.6 Luna, the ChatGPT Chat models: February 16, 2026.
• Claude Haiku 4.5: February 2025. Gemini 3.1 Pro: January 2025.
• DeepSeek V4.1 Flash and V4 Pro: not stated on the DeepSeek pages we read.

The detail most people miss: the default chat models in ChatGPT stop in February 2026, the oldest of the three big assistants. Claude's free default knows the world to June 2026 and Gemini's Flash models to March. Search inside the apps narrows the gap but does not close it, because a model that has read a page can still misstate it.

For SEO content the stakes are the ranking itself. Google Search Central says it focuses on the quality of content rather than how it was produced, and warns that publishing many AI pages without adding value may violate its scaled content abuse policy. An Ahrefs study published in July 2026 found that pages under 50% AI content held 82.2% of top-three rankings; there is more of that data in our AI SEO statistics roundup. The fix is the same on every model: put the facts in the brief with a source for each, then have the model list every factual claim in its draft with the sentence that supports it, and open each source. That check is the whole job of an AI fact checker for drafts, and it is the step most chat workflows skip.

A worked example: one month of content, five models.

Take a fictional bookkeeping software company, Tallyroom, running a monthly program: ten blog articles in English, four more written in Spanish for its customers in Mexico, and a new title and meta description for each of 400 help-center and product pages. The company and everything it logged are invented. The prices, cutoffs and token math are the real ones from the sections above, at list price before thinking tokens.

API spend for the month: under $3, plus the ChatGPT plan. Human time: about 28 hours, spent on briefs, the Spanish edit, the claim check and publishing. That ratio is the real answer to which LLM is best for writing. The model choice is worth getting right because it is cheap to get right, and then nearly every remaining hour sits in work no model does. If the program also has to get Tallyroom quoted when people ask an assistant about bookkeeping software, that is a separate job, covered under LLM SEO for AI answers.

Common mistakes when picking an LLM for writing.

1. Picking from one leaderboard. A single number is easy to quote, so it gets quoted. GPT-6 Astra is 4th on short stories and 11th on editorial voice: each board tells you about its own job. Use boards to build a shortlist, then decide on your own brief.
2. Using one model for every job. It saves a decision, which is why it is the default. Claude Fable 5.1's output costs a hundred times GPT-6 Luna's, and on 800 meta descriptions that buys nothing a reader can see. Route each job to the cheapest model that does it well.
3. Judging a model you are not actually using. Launch posts describe the model; the app gives you whatever your plan includes. GPT-6.1 Sol is not in ChatGPT's Chat, Gemini 3.8 Flash needs Google AI Pro in the app, and Claude Opus 5.5 needs a paid Claude plan. Check the model menu before you judge the writing.
4. Asking the model for search volume. It answers confidently, so the number feels real. No model has a keyword database; take the query from measured data or your own Search Console before any prompt.
5. Translating instead of writing. An English draft run through a model is the fastest way to a Spanish page, and it reads like one. Brief in the target language, write natively and have a native speaker edit.
6. Upgrading the model to fix a volume problem. When output stalls at a few articles a week, a better model is the tempting fix. The bottleneck is checking time per article, and it scales with people, not models, which is the case for autoblogging that stops weak drafts before they publish.