Which tool to choose
Three free tools cover almost all cases. The choice depends on how many variants you need and how long they are.
If you need many short variants in little time (titles, email subject lines, social captions, slogans): use ChatGPT on the free plan. It produces long numbered lists without trouble, but watch out for the limit: the free plan gives about 10 messages every 5 hours on the GPT-5.5 Instant model, then switches to a faster mini version. That's why it's worth asking for all the variants in a single message.
If you have to vary long texts while keeping coherence and nuance (articles, articulated product descriptions, complex emails): use Claude. The free plan runs on Claude Sonnet 4.6, while Opus 4.8, the most powerful model for demanding tasks, is reserved for Pro and Max subscribers. Sonnet 4.6 is more than enough for rewriting, and the limit is dynamic, generally between 20 and 50 messages every 5 hours: you have more room to iterate sentence by sentence.
If you want everything inside Docs and Gmail, or you're looking for an alternative when the other two have exhausted the limit: use Gemini. The free plan allows up to 30 prompts a day, enough for variant sessions scattered through the day.
Practical tip: keep two tabs open. When one hits the cap, you switch to the other without losing the thread.
How to do it
The principle: the AI produces useful variants only if you give it an axis of variation, that is, the dimension along which the versions must differ. "Give me some alternatives" produces five almost identical sentences. "Give me five versions that change the tone" produces genuinely different material.
From browser or app, the path is the same:
- Open a new conversation on ChatGPT, Claude or Gemini.
- Paste the original text between quotes or after a colon, so the AI understands where your instruction ends and the material to work on begins.
- Specify the number of variants, the axis of variation and the constraints (maximum length, words to avoid, audience).
- Ask for numbered output with a label for each version: you need it to say "develop number 3 better for me".
- Once you've read the list, choose one and ask for a second round on that one only ("From version 3 give me 4 more variations changing only the opening").
The syntax for multiple variants in one shot:
Rewrite the text below in 6 numbered variants. Keep the central message unchanged, but change the axis indicated for each:
1) formal and institutional tone
2) colloquial tone, as if talking to a friend
3) short version, 30 words max
4) version that leverages emotion and desire
5) version that leverages concrete data and measurable benefit
6) version with a direct question at the opening
Constraints: no inflated superlatives, no multiple exclamation marks, adult English-speaking audience.
For each variant add a one-word label that sums it up.
Text to rewrite:
"[paste your text here]"
Check: if the six versions seem interchangeable, the axis of variation was too weak. Add more extreme constraints ("version 3 must fit in a text message") and try again.
To polish the winning version inside the text instead of regenerating the whole chat, ChatGPT has Canvas, the side editor where you edit the document. A note: GPT-5.5 Instant supports ChatGPT's tools except Canvas; if you need to create or edit in Canvas, select GPT-5.5 Thinking. On Claude the equivalent feature is Artifacts, available on the free plan too.
Concrete example
Marta manages the Instagram page of a pastry shop. She needs to announce the artisanal panettoni and has written: "Our handmade panettoni have arrived. Come and try them."
She pastes the sentence into Claude and uses the prompt above asking for 6 variants for Instagram, 25-word constraint, local audience.
Result in a few seconds: the "emotion" version talks about the scent that fills the workshop at five in the morning; the "concrete data" one cites the 48 hours of natural leavening; the "direct question" opens with "Have you already decided on the dessert for the 25th?". Marta discards the first two (too generic), keeps the 48 hours, and asks for a second round: "From the version on the 48 hours give me 3 different openings, one must contain a single emoji." She chooses the opening with the clock, adds the price by hand and publishes. Total time: four minutes, against the half hour usually spent staring at the cursor.
Variants aren't there to publish them all, but to make you choose faster by discarding the weak one.
When it does NOT work (and how to fix it)
If the variants are all the same
It means you asked for "some alternatives" without a criterion. Rewrite the instruction imposing axes opposite to each other: one short and one long, one rational and one emotional, one that opens with a figure and one with a question. The more distant the axes, the more the versions diverge.
If you've used up your messages mid-work
On free ChatGPT, once you hit the cap the chats switch to the mini version of the model until the limit reset. The mini still handles simple variants: continue there, or move the session to Gemini or Claude in the other tab. To not waste messages, always ask for the complete block of variants in a single request instead of one at a time.
If the AI changes the meaning instead of the style
This happens when the original text is ambiguous. Add a line before the prompt:
The message to preserve in all versions is exactly this: [write in one sentence what must NOT change]
This way you constrain the content and let only the form vary.
If you need to edit the text, not regenerate it from scratch
Regenerating in chat makes you lose your manual edits at every round. Switch to Canvas on ChatGPT (remembering you need GPT-5.5 Thinking) or to Artifacts on Claude: there you select the sentence to change and the AI touches up only that one, leaving the rest intact.
A tip from someone who actually uses it
Don't ask for a low number of variants thinking you'll choose half. Ask for 7 or 9: the AI, forced to fill the list, has to get out of the first two or three obvious ideas and arrives at lateral solutions, often the most interesting ones. The first three variants of any request are predictable; the value appears from the fourth on, when the model has exhausted the obvious answers.
Frequently asked questions
Can I have variants created starting from a very long text?
Yes, and Claude is the most suitable: the free plan runs on Sonnet 4.6 with a 200K-token context window, enough to load and discuss long documents. For huge texts it's still worth dividing into sections and asking for variants section by section, so you control the result better.
How many variants can I get without paying?
It depends on how you spend the messages. By asking for all the variants in a single request, even ChatGPT's free plan with its ~10 messages every 5 hours is enough for hundreds of alternatives a day. The limit counts messages, not output lines: a single message can contain nine of them.
Are the generated variants original or do I risk copying something?
The variants are rewrites of the text you provide, so they start from your content: plagiarism isn't the problem. The problem is homogenization. Used by thousands of people from the same models, "AI" phrases start to resemble each other across different brands, and people who read a lot online recognize them at a glance. That's why the last step is yours: take the best variant and change by hand at least the opening and one characterizing word. The AI takes you to 90%, the last 10% that sets you apart you put in yourself.