Best ChatGPT model for writing: GPT-5.6 Sol, GPT-6.1 Sol, GPT-6 Astra and Luna compared.

1. The current ChatGPT models, and where you get each.

“ChatGPT” is an app, not a model, and since July 2026 the app has two places to write. Chat is the conversation you know. Work is an agent that returns finished documents, sheets and slides. The two run different model families: the chat window runs GPT-5.6 Sol and GPT-5.6 Luna, while the GPT-6 models released in September live mostly in Work, in Codex and in the API. So the first question is not which model is best, but where you are writing.

Here is every current model a writer can reach as of September 30, 2026, with where you get it, the writing job it suits, its API price per million input and output tokens, and its knowledge cutoff from ChatGPT's API docs.

Three details the table cannot hold. Every model on it reads about 1.05 million tokens at once and writes up to 128,000, so length is never the constraint for an article. GPT-5.6 Sol's $4 and $20 is promotional pricing that runs at least to November 21, 2026. And GPT-6 Astra shows up in the chat window only under the name GPT-6 Pro, only on the Pro, Business and Enterprise plans, and with a weekly message cap; Plus users reach Astra in Work and Codex instead, on an allowance that resets every five hours.

If you are still choosing between assistants rather than models, start with our Claude vs ChatGPT for writing comparison, which matches the tiers fairly and gives you a same-brief test. If your team writes in Claude instead, the equivalent pick per job is in our guide to the best Claude model for writing.

2. What changed in ChatGPT models (September 2026).

September was the busiest month for ChatGPT models this year. GPT-6 Astra arrived on September 4, GPT-6 Sol and GPT-6 Luna on September 22, and GPT-6.1 Sol a week later on September 29, as an upgrade to GPT-6 Sol. None of the four replaced the models in the chat window. What that means for a writer:

• The best new writer is not in Chat. GPT-6.1 Sol runs in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and in the API. Its launch coverage says it is not yet available in Chat.
• Cheaper per word. GPT-6.1 Sol costs $2 and $10 per million input and output tokens, a fifth of GPT-6 Astra's price. The September 22 launch priced GPT-6 Sol and Luna at half the GPT-5.6 series.
• Fewer wrong facts. GPT-6.1 Sol's launch reported its factual error rate falling from 11.4% to 7.7% against GPT-6 Sol at low reasoning effort. Fewer is not none (section seven).
• Newer knowledge. The GPT-6 models know the world to April 30, 2026 (Astra, 6.1 Sol) or May 18, 2026 (Luna). The GPT-5.6 models behind the chat window stop at February 16, 2026.
• A new place for long drafts. Pages, a shared document type for people and agents, launched on September 29 for the Pro, Business and Enterprise plans.

How the lineup got here:

1. 2026-05-28 · Canvas removed from GPT-5.5; drafts move into editable writing blocks inside replies.
2. 2026-07-09 · GPT-5.6 released to the public, and ChatGPT Work launched with it.
3. 2026-08-06 · GPT-5.6 Luna becomes the Free and Go default with unlimited text chats; GPT-5.6 Sol gets the thinking slider on Plus and Pro.
4. 2026-09-04 · GPT-6 Astra, after a one-day preview.
5. 2026-09-22 · GPT-6 Sol and GPT-6 Luna, in Work, Codex and the API.
6. 2026-09-29 · GPT-6.1 Sol and Pages.
7. NEXT · GPT-6.1 Sol in the chat window, no date given.

We update this page when a ChatGPT model ships or moves between surfaces. It was last checked against ChatGPT's API docs and launch coverage on September 30, 2026. The version-free companion, which covers the SEO setups rather than the models, is our ChatGPT for SEO guide.

3. The best ChatGPT model for writing in the chat window.

Most writing still happens in the chat window, and there the choice is narrower than the headlines suggest. On Plus you write with GPT-5.6 Sol. On Free and Go you write with GPT-5.6 Luna. On Pro, Business and Enterprise you get GPT-5.6 Sol plus GPT-6 Pro, which is Astra under a weekly cap. For most people, then, the question is how to use Sol well, and the answer is the thinking slider.

• GPT-5.6 Sol, slider low. Emails, social posts, short rewrites, headline options. Fast, and its August update made answers more direct and cut unnecessary formatting, which is most of what makes chat output read like chat output.
• GPT-5.6 Sol, slider high. Articles and anything that has to follow a detailed brief. ChatGPT's own description of the slider says higher settings are meant for planning, research, writing and coding. It answers more slowly; for a 2,000-word draft that is worth it.
• GPT-5.6 Luna with Think. The free option, with unlimited text chats. Think gives Luna more time on a harder request, but it does not switch you to a stronger model, so long, voice-led pieces need more editing than on Sol.
• GPT-6 Pro (Astra). Pro plans and up only, capped weekly. Worth it for fiction and research-heavy reports; not the upgrade it looks like for articles in a house voice (section four).

Two settings matter as much as the model. Drafts now land in editable writing blocks inside the reply, since ChatGPT removed Canvas from its current models in May 2026, so edit there and ask for changes one section at a time rather than regenerating the whole piece. And put your style rules, banned phrases and internal link list in a Project: project instructions apply only inside that project and override your global custom instructions, which is exactly what you want for one site's voice. The full setup is in how to set up a ChatGPT Project for SEO writing.

4. GPT-6 Astra vs GPT-6.1 Sol for writing.

If you write in ChatGPT Work or through the API, the choice is between the flagship and the model that nearly matches it. ChatGPT's API docs tell developers who are unsure to start with GPT-6 Astra, and describe it as the flagship for complex reasoning and coding. Neither of those is writing, and the early evidence on Astra's prose splits by the kind of writing.

On fiction, Astra leads GPT-5.6 Sol. It entered Lech Mazur's Short-Story Creative Writing Benchmark, which scores complete short stories written to the same constrained briefs, at number three, and beat GPT-5.6 Sol head to head. On house-voice nonfiction it trails. A Decrypt review published on September 6, 2026 reported that in Louis-François Bouchard's benchmark for writing in an editorial voice, Astra ranked 11th, below GPT-5.6 Sol in 6th, a result he called surprisingly disappointing, and other testers found its prose flat. The same review had Astra excelling on work with a verifiable answer. Neither benchmark has scored GPT-6.1 Sol yet, so the comparison below rests on its specs and its launch claims.

• Pick GPT-6.1 Sol for articles, blog posts, rewrites and finished documents in Work. Its model page describes near-Astra performance for professional work, it costs a fifth as much per token, and it shares Astra's April 30, 2026 knowledge cutoff.
• Pick GPT-6 Astra for fiction, and for one long document built from a large research pack, where a wrong number costs more than a flat sentence. Test it against GPT-6.1 Sol on the same material before you decide.
• The cost gap per article. A 3,000-word brief in and a 2,000-word article out is about 4,000 and 2,700 tokens, using the help center rule of thumb of 100 tokens per 75 English words. At list price that is about 3.5 cents on GPT-6.1 Sol and about 18 cents on Astra, before reasoning tokens.

Per article, both are cheap next to the edit. The reason to prefer 6.1 Sol is not the 15 cents; it is that, for articles in a house voice, the evidence so far does not show Astra writing better prose for the money. If you are shopping for tools that wrap these models for search work rather than calling them yourself, our list of ChatGPT SEO tools covers the field.

5. The best ChatGPT model for creative writing.

ChatGPT's model docs do not rank the models for creative writing, and public leaderboards disagree depending on which board and which version you read. So treat any confident list, this one included, as a starting point for your own test. With that said, here is how the current models line up for fiction, scripts and other creative work.

• GPT-6 Astra for fiction, if you have it. It entered Lech Mazur's short-story benchmark at number three, ahead of GPT-5.6 Sol, the strongest public fiction result we found for a current ChatGPT model. In the chat window it is GPT-6 Pro on the Pro, Business and Enterprise plans; Plus users reach it in Work. Some testers still found its prose flat, so judge it on your own scenes.
• GPT-5.6 Sol is the everyday default. The best writer you can pick in the chat window on Plus, and the slider gives you room on long scenes. Raise it for chapters, keep it low for brainstorming.
• GPT-6.1 Sol for long projects in Work. A series bible, a manuscript with its notes, or a nonfiction book from a research pack. It has newer knowledge than GPT-5.6 Sol and holds a large pack in one go.
• The Luna models are for volume. Twenty premises, alternate lines of dialogue, name ideas. Not finished prose.

What decides creative quality is the sample you give the model. Paste 500 to 1,000 words of your own finished prose as the style reference, list the words and moves you never use, and ask for three versions of an opening rather than one. Then do the line edit yourself, because that is where a draft becomes yours. Commercial creative work (taglines, ad copy, landing page headlines) is a different job: mostly generating options for a person to pick from, which the Luna models do cheaply. If you want tools built around that job, see our roundup of AI copywriting tools.

6. The best ChatGPT model for SEO content and blog posts.

For articles meant to rank: GPT-5.6 Sol on a high thinking setting if you draft in the chat window, GPT-6.1 Sol if you draft in Work or through the API, and GPT-6 Luna for titles and meta descriptions across many pages. Past that choice, the model is the least important input to an article that ranks.

Three things decide it, and no ChatGPT model supplies any of them. There is no keyword database inside the model, so any search volume it quotes is a guess; the query has to come from measured data or your own Search Console, as our guide to blog SEO explains. It has never seen your customers, your numbers or your product, so the first-hand material that makes a page worth ranking has to come from you. And every model states outdated or wrong facts with confidence, so the claims need a check (section seven).

Google is clear that the tool is not the issue. Google Search Central says it focuses on the quality of content rather than how it was produced, and warns that using generative AI to publish many pages without adding value may violate its scaled content abuse policy. The ranking data points the same way: an Ahrefs study published in July 2026 found that pages under 50% AI content held 82.2% of top-three rankings, and there is more of that data in our AI SEO statistics roundup. The human share of the page is what separates it, not which Sol typed the first draft.

So spend the effort where it pays. Write the brief before any draft prompt (the process is in our guide to ChatGPT for content creation), draft one section at a time, and keep the brief, the sources and the style rules in one Project. If the goal is also to be the page ChatGPT quotes when someone asks it a question, that is a separate discipline, covered under LLM SEO for getting cited in AI answers.

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

Every ChatGPT model has a knowledge cutoff, the date its training data stops. For writing it is the number that matters most after price, because anything newer is either missing or guessed. As of September 30, 2026:

• GPT-6 Luna: May 18, 2026, the most recent of any current model.
• GPT-6 Astra and GPT-6.1 Sol: April 30, 2026, five months before GPT-6.1 Sol shipped.
• GPT-5.6 Sol and GPT-5.6 Luna: February 16, 2026. These are the models behind the chat window, so most ChatGPT writing today starts seven months out of date.

Web search in ChatGPT narrows the gap, and ChatGPT's help center is candid that search citations can be incomplete, outdated or incorrect. The error rates are falling: GPT-5.6 Sol's August update reported about 68% fewer factual errors than GPT-5.5 Instant, and GPT-6.1 Sol cut GPT-6 Sol's rate from 11.4% to 7.7% at low effort. Lower is not zero, and an article carries dozens of claims.

The fix is the same on every model. Put the facts in the brief yourself, with the source for each. After the draft, ask the model to list every factual claim with the source it used and the sentence that supports it, then open each source and cut whatever has none. That check is the whole job of an AI fact checker, and it is the step most chat-window workflows skip.

A worked example: one content program, four models.

Take a fictional garden-supply store, Loamwell, running a monthly program: twelve blog articles, one long buying guide, and new meta descriptions for 150 product and category pages. The store and everything it logged are invented. The model prices, cutoffs and token math are the real ones from the sections above, before reasoning tokens.

For reference, here is what one standard article (the 3,000-word brief in, 2,000 words out) costs on each model through the API at list price:

Loamwell's model spend for the month was under $2 on top of the Plus seat, and reasoning tokens, which bill as output, add to that without changing the order of magnitude. The human side was about 30 hours: briefs, the claim check, the edit and publishing. That ratio is the real answer to which ChatGPT model is best for writing. Pick the right model per job, as above, then look hard at the 30 hours, because that is where both the cost and the quality live. At volume, short copy is also worth comparing against a purpose-built meta description generator or a free SEO title generator that checks length against what Google shows.

Common mistakes when picking a ChatGPT model.

1. Assuming the newest model is in Chat. Launch headlines name GPT-6 models; the chat window on Plus still runs GPT-5.6 Sol. Check which surface a model is on before you judge it, or you will compare a model you never used.
2. Paying for the flagship to write house-voice articles. GPT-6 Astra is the safe-sounding choice and the API docs' default for complex reasoning. Its benchmark win is on fiction; on editorial-voice writing the early reviews put it behind GPT-5.6 Sol, at five times the per-token price of GPT-6.1 Sol.
3. Leaving the slider at Instant. It is the fast default and fine for an email. For an article that has to follow a detailed brief, ChatGPT says the higher settings are the ones meant for writing.
4. Trusting what the model remembers. A fluent sentence reads as a checked one. The chat models stop at February 16, 2026; put the facts in the brief and check whatever came from the model.
5. Judging models on a blank prompt. “Write a blog post about X” in two windows returns two averages of the web. Run your real brief and style sample through both, count the edits, and keep your rules in the prompts and Project instructions you reuse.
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 gates quality before it publishes.