Today we’re launching Superstealth, our newest and strongest humanizer model. At the same time, we’re retiring our legacy v3 models. Here’s what changes for you, and what you don’t need to change.
Rephrasy Team
Oct 20, 2026
Superstealth is built for one goal: the highest bypass rates against modern AI detectors. Instead of swapping words, it rewrites text the way a person would. Sentence rhythm, structure and phrasing all change, while the meaning stays the same.
It takes a bit longer. A typical text of a few hundred words is ready in 10 to 30 seconds.
Long texts can take up to a minute. The extra time goes into a deeper rewrite.
Length stays close to your original. Output is usually about the same length as the input.
Superstealth is rolling out on the AI Humanizer now. If it’s available for you, you’ll find it at the top of the model selector with a Beta badge, already preselected.
Superstealth is live in the Rephrasy API. You only need to change the model name:
Responses come back in the same format as every other model. Because Superstealth works harder on each text, set your client timeout to at least 120 seconds.
With Superstealth and v4 in place, we’re retiring our legacy v3 models. You don’t need to change any code. Existing requests are routed automatically:
You send | You get now |
|---|---|
v3 via the API (all styles) | Superstealth |
v3 via the API with HTML or Markdown input | v4 |
The response format stays the same. The style option (creative, journalistic, professional) was specific to v3 and is no longer applied.
If you call v3 through the API: your requests now run on Superstealth. Please make sure your client timeout is at least 120 seconds.
Not affected: v4, ft_clean-v3, Undetectable Model, SEO Model and your own Writing Styles keep working exactly as before.
Superstealth: when bypass rate matters most.
v4: fast and reliable, our recommended default for most texts.
ft_clean-v3: when meaning and length must stay as close to the original as possible, for example technical text with figures.
Superstealth stays new and we might change some small setting to improve the quality even further in the next couple of days.
We tested Superstealth on 88 fresh texts written by current AI models (Gemini 3.5 Flash, GPT-5.4 and
GPT-5.4-mini): student essays and lab reports, academic paper introductions and discussions, blog and
Substack posts, op-eds and news articles. Every text went through Superstealth once, exactly as it would
through the website or the API, and was then checked by three detectors.
Pangram is the strictest detector on the market today, and the one most humanizers fail completely. Almost
half of the texts pass it on the first try. Because every Superstealth run writes the text differently, a
second run on the texts Pangram still flagged brought the share to + 60%. If a text doesn't pass, simply
run it again.
GPTZero and ZeroGPT rate three out of four and nine out of ten Superstealth texts as human-written.
Meaning stays intact. Every rewrite is checked for meaning, grammar and fluency before you see it, and
the text keeps its length (usually within ±10%).
Detectors change often. These results were measured in October 2026; we will publish new numbers as the
detectors and Superstealth evolve.
Convert AI to Human Text — Free AI Humanizer
The #1 AI humanizer trusted by 200,000+ students & professionals