If you mostly use ChatGPT to write, the alternatives worth trying are Claude for long fiction and polished prose, Hemmingway for emails, messages and anything that has to sound like you wrote it, GLM-5.3 and Kimi K3 for stories, and Hemmingway-1 if you want a model with open weights. None of them is best at everything, so the right pick depends on what you write.
We build a writing model of our own, Hemmingway-1, so we test the big models against each other all the time. The numbers on this page come from those tests, run in September 2026. We say which models we tested and which we did not, and where our own model loses.
The models we tested
For ChatGPT, the model in our tests is GPT-6 Astra. For Claude, it is Fable 5.1 and Fable 5, with Opus 4.7 and Opus 4.8 on the public EQ-Bench 4. We also tested Kimi K3 and GLM-5.3. Hemmingway-1 is ours.
| Model | Everyday messages | Sounds like a person | Long stories | EQ-Bench 4 |
|---|---|---|---|---|
| Hemmingway-1 | 1026 | 1032 | 1197 | 1330 |
| Fable 5.1 (Claude) | 1024 | 1006 | 1350 | not run |
| Fable 5 (Claude) | 1009 | 992 | 1277 (Fable 5 Max) | 1341 |
| GLM-5.3 | 1007 | 996 | 1254 | not run |
| Kimi K3 | 996 | 985 | 1197 | 1332 |
| GPT-6 Astra (ChatGPT) | 976 | 964 | 1312 | GPT-5.5: 1316 |
Everyday messages is CommunicationBench, sounds like a person is Human-Likeness, and long stories is StoryBench. All three are our own benchmarks: blind, run in both orders, with a judge that is not one of the models being judged. EQ-Bench 4 is a public benchmark that we ran ourselves with its official harness. It compares models in its own snapshot, so ChatGPT is represented there by GPT-5.5. All scores are on an Elo scale, where higher wins more head-to-head matchups.
Best for long fiction: Claude
If you write novels, long chapters or anything that has to hold a story together over many pages, Claude is the strongest alternative we have tested. Fable 5.1 scored 1350 on StoryBench, the highest of any model on our board, and beat GPT-6 Astra there. Fable 5 also placed first on EQ-Bench 4 in our comparison, which matters for fiction that turns on how characters feel.
The catch for short writing: Claude's models tend to wrap the message. Fable 5 put its answer inside commentary, options or notes in more than nine replies out of ten in our tests, and Fable 5.1 did it about two times in three. The writing is very good; you just have to cut it out before you use it.
Suits: novelists, long-form fiction, anyone who wants the most capable general assistant that also writes well.
Best for emails, messages and sounding like you: Hemmingway
Hemmingway is built for the writing people do every day: the reply to a client, the text to a landlord, the email you keep putting off. Its model, Hemmingway-1, came first of the models we tested on everyday messages, level with Fable 5.1 and fifty points ahead of GPT-6 Astra. It also came first on sounding like a person, twenty-six points clear of the next model.
The gap is widest on hard asks: saying no, asking for money back, telling someone something they will not like. Judged on which reply sounded like a person wrote it, GPT-6 Astra's won 9% of the time and Hemmingway-1's 72%. And it gives you one message, not a menu of options with notes attached.
The apps for Mac, Windows and Android can read your mail and chats and draft each reply for you to send. You can tell it who people are, and it writes to each of them the way you would. It is also available in the browser at /app/.
Where it loses: long fiction. Claude, GLM-5.3 and GPT-6 Astra all write better long stories in our tests. It is a writing tool, not a general assistant for code or research. It is English-first, and it can be wrong and still sound certain.
Suits: people who write a lot of email and messages, anyone who wants the words to sound like theirs, and people who want a model with open weights. More on how it compares with ChatGPT.
Also strong on stories: GLM-5.3 and Kimi K3
Outside ChatGPT and Claude, GLM-5.3 and Kimi K3 are the other two story writers we tested. GLM-5.3 is the stronger story writer of the two in our tests, at 1254 on StoryBench, behind only Fable 5.1, GPT-6 Astra and Fable 5 Max. Kimi K3 sits at 1197, level with Hemmingway-1, and scored 1332 on EQ-Bench 4, just behind Fable 5.
On everyday messages both came in behind Hemmingway-1 and Fable 5.1. Like Fable 5, both put the message inside commentary, options or notes in more than nine replies out of ten, so expect to edit.
Suits: writers who want a strong story model outside the two big assistants. If you want a model with open weights for messages and short writing, Hemmingway-1 is under Apache-2.0. More on open models for writing.
When ChatGPT is still the right choice
To be fair to ChatGPT: GPT-6 Astra's messages were short, about 59 words on average, and it almost never wrapped them in commentary. What you get is usable as it is. It was also the second-best story writer we tested, behind Fable 5.1. Where it fell behind was in sounding like a person rather than a careful assistant.
If you use one tool for writing, code, research and questions, ChatGPT is still a sensible default. Many people keep a general assistant for thinking and a writing tool for the words that go out under their name.
Models we did not test
Google's Gemini, Jasper, Sudowrite and many other writing tools are not in our results, so we cannot rank them here. If a model is missing from the table above, it is because we did not test it, not because we tested it and it did badly.
Try HemmingwayDownload the app
The ranking by use case
| If you mostly write | Try first | Also good |
|---|---|---|
| Emails, texts, replies | Hemmingway | Fable 5.1 |
| Messages that must sound like you | Hemmingway | Fable 5.1 |
| Hard asks: no, refunds, bad news | Hemmingway | Fable 5.1 |
| Novels and long chapters | Claude (Fable 5.1) | GPT-6 Astra, GLM-5.3 |
| Stories, outside ChatGPT and Claude | GLM-5.3 | Kimi K3, Hemmingway-1 |
| Short messages to paste as they are | Hemmingway | GPT-6 Astra |
| Writing plus code and research | ChatGPT or Claude |
How to try an alternative properly
- Pick three things you actually wrote last week: an email, a text, and something longer.
- Give each tool the same facts and the same request. Do not tidy the request for one and not the others.
- Ask for "just the message, one version, no notes". This shows you which tools follow the instruction.
- Read each answer aloud. Count how much you would have to cut or change before sending it.
- Keep the one you edited least for that kind of writing. It is fine to keep two.
All of Hemmingway-1's scores are on the model card.
Common questions
What is the best alternative to ChatGPT for writing?
It depends on what you write. In our tests Claude's Fable 5.1 wrote the best long stories, and Hemmingway-1 came first on everyday messages and on sounding like a person. For fiction try Claude; for email and messages try Hemmingway.
Is Claude better than ChatGPT for writing?
In our blind tests Claude came out ahead of ChatGPT on every writing measure we ran: everyday messages, sounding like a person, long stories and EQ-Bench 4. ChatGPT's GPT-6 Astra was more direct, and rarely wrapped its messages in commentary. The full comparison is here.
Is there a free ChatGPT alternative for writing?
Hemmingway has a free trial with no card needed, in the browser and in its apps. Its model, Hemmingway-1, is also open weights under Apache-2.0, so you can run it yourself. For what paid plans cost, see /pricing/.
Which open-source model is best for writing?
Of the models in our tests, Hemmingway-1 is the one we know has open weights, under Apache-2.0. It came first on everyday messages and sounding like a person. For long stories, GLM-5.3 scored higher than it on our StoryBench.
Did you test Gemini?
Not for these results. Gemini is not in our test tables, so we do not rank it on this page.