A guide from ZeroTwo
Which model writes best depends on the job and on your ear. This page gives you a shortlist for eight writing jobs and a blind test you can run in an afternoon, and it is straight about what nobody, including us, has measured.
Short answer
No model wins every writing job, and we have not run a test that ranks them. Shortlist two or three models for your job, give them the same brief, and judge the outputs blind.
Disclosure
ZeroTwo publishes this guide and sells access to many of these models. That is the reason it recommends testing over trusting a ranking from us.
These are starting candidates, not winners. Each row says what to judge first, and lists models in no particular order. Add or drop models freely; the test decides.
| Writing job | What to judge first | Candidates to test (unranked) |
|---|---|---|
| Long-form blog posts | Voice and argument past ~1,000 words; cleanup time | Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro |
| Marketing copy and ads | Brand-voice fit; claims stay inside the brief; usable variants | GPT-5, Claude Sonnet 4.6, Grok 4 |
| Marketing emails | Subject-line variety; tone; no invented offers | GPT-5, Claude Sonnet 4.6 |
| Fiction and creative writing | Character voice; continuity across scenes; phrasing that is not generic | Claude Sonnet 4.6, GPT-5, Grok 4 |
| Screenplays and scripts | Dialogue that sounds spoken; structure; continuity | Claude Sonnet 4.6, GPT-5 |
| Academic and essay writing | Argument structure; every reference real and accurately described | Gemini 3.1 Pro, Claude Sonnet 4.6, GPT-5 |
| Technical documentation | Terminology precision; consistent structure; working examples | Claude Sonnet 4.6, DeepSeek Chat, GPT-5 |
| Social media posts | Tone per platform; freshness; claims you can back up | Grok 4, GPT-5, Claude Sonnet 4.6 |
Editorial starting points chosen to cover different vendors. No benchmark, internal test or rater panel stands behind the order or the selection.
Three situations, three different answers. In none of them is the answer a model name.
A default chosen on your own brief beats re-reading comparisons. Re-test when your model changes.
Any model can invent a reference. Budget time to open each source, whichever model you pick.
A real paragraph of yours and one-change-at-a-time edits move voice more than switching models does.
Once you have a default, the work moves from choosing to writing. That loop, from brief to revision to fact-check, is on the AI writing assistant page, where you can draft and revise in the writing workspace.
Six messages cover two briefs on three models. The Free plan starts with 100 free credits and adds 20 bonus credits every day you log in, though it offers a limited model selection, so some candidates may need a paid plan.
Audience, purpose, length, a paragraph of your own writing as the voice target, and the facts the piece may use. Use a job you actually have, not a toy prompt.
Same brief, same sitting, same instructions. Do not tweak the prompt between models, or you are testing your edits instead of the models.
Paste each output into a document and label them A, B and C in shuffled order. Ideally someone else does the labelling, so you cannot recognise a favourite.
Score each output on the four criteria below before you look at the labels. Write the cleanup time down honestly; it is the number that decides daily use.
Run a second brief of the same type. If the same model wins twice, make it your default for that job. Re-test when you switch models or when a new version ships.
Four criteria, scored before you unblind the labels. Copy the scorecard into a note and fill it in per brief.
| Criterion | A | B | C |
|---|---|---|---|
| Voice match | blank | blank | blank |
| Factual safety | blank | blank | blank |
| Cleanup effort | blank | blank | blank |
| Fit for the reader | blank | blank | blank |
We do not restate specs or prices here, because they change. Each card links to that model's page. The “look for” line is a question for your test, not a verdict.
Anthropic
Include it in tests for
Long-form, fiction, scripts, documentation
Look for
Does the voice and the argument hold past the first thousand words? How much do you cut?
Claude Sonnet 4.6 model pageOpenAI
Include it in tests for
Marketing copy, emails, format-heavy pieces
Look for
Can it hold your brand voice across variants without drifting into generic copy?
GPT-5 model pageInclude it in tests for
Source-heavy and academic drafting
Look for
Does every reference it gives exist, and does it say what the draft claims?
Gemini 3.1 Pro model pagexAI
Include it in tests for
Short social posts and casual tone
Look for
Is the tone right for the platform, and does it stay inside the facts you gave it?
Grok 4 model pageDeepSeek
Include it in tests for
Technical documentation
Look for
Is the terminology consistent, and do the examples actually run?
DeepSeek Chat model pageTwo well-known studies of generative AI at work. Neither tests one model against another, so neither can tell you which model to use for a writing job. Both abstracts were read on their publishers' pages in October 2026.
What it shows
In a preregistered online experiment with 453 college-educated professionals, the half given ChatGPT took 40% less time on mid-level professional writing tasks, and output quality rose 18%. Weaker writers benefited most.
What it does not show
It used one tool, ChatGPT, so it says nothing about which model suits which writing job. The tasks were incentivised professional writing exercises, not fiction, long-form or publication-grade work.
What it shows
Across 5,179 customer-support agents, access to a generative AI assistant raised issues resolved per hour by 14% on average, and by 34% for novice and low-skilled workers.
What it does not show
It studies customer-support conversations, not writing, and compares workers with and without one assistant, not models with each other.
There is no single best AI for writing, and we have not run a test that would let us name one. The right model depends on the job and on your ear. Shortlist two or three candidates for your job from the table on this page, run the same brief through each, and judge the outputs blind.
We cannot give a verified winner, and anyone who does is usually describing a taste. Include both in your test on the same brief and compare voice match, unsupported claims and cleanup time. The ChatGPT vs Claude comparison on ZeroTwo covers the differences between the products.
Start from the shortlist for that job and test it. For fiction, judge character voice and how generic the phrasing is. For academic work, judge whether every reference exists and says what the draft claims; that check matters more than which model wrote the draft.
Treat output as a draft. Fact-checking, voice and originality still need a human pass. The published research is about speed and quality on specific tasks: in one preregistered experiment with 453 professionals, ChatGPT users took 40% less time on mid-level writing tasks and produced work rated 18% higher. That study did not test publication standards.
Use them to choose candidates, not winners. A leaderboard measures a particular set of prompts and raters, and rarely measures your voice, your reader or your cleanup time.
Yes. ZeroTwo's Free plan costs $0 and includes 100 free credits to start plus 20 bonus credits every day you log in, one file upload per message and a limited selection of models and tools, with no card required. Which models are included on Free is shown in the app, so some candidates may need Plus ($14.99 per month) or Pro ($29.99 per month).
ZeroTwo gives you 60+ models in one app, so you can run the same brief on several of them without juggling accounts. ZeroTwo sells that access, which is why this page tells you to test models yourself rather than trust a ranking from us.
The draft-and-revise workflow once you have a default model.
From keyword to fact-checked post.
Structure and clarity for essays; verify every citation.
Editing a long document you already have.
Choosing models when sources matter most.
An editing interface for work that needs revisions.
How the two products differ.
The model catalog, with a page for each model.