Blog AI Humanizer

Rephrasy AI Humanizer Review: 10 of 50 Pangram Passes

Rephrasy Superstealth passed 10 of 50 Pangram tests and zero Originality.ai tests in the November 2026 HumanizerEval benchmark. StealthGPT Super passed 44 of 50 on Pangram and 19 of 50 on Originality.ai on the same English drafts. Rephrasy’s best result was 44 of 50 on Winston AI. It also finished fourth overall among 15 humanizers, with a 37.49 out of 100 combined score. StealthGPT Super placed second at 70.34. See the November leaderboard.

The benchmark tested Rephrasy’s Superstealth API and StealthGPT’s Super API in Quality mode with rephrasing enabled. StealthGPT’s team runs HumanizerEval and StealthGPT is one of the products it tests. Below are the results for all five detectors, what the quality checks say about the rewrites and where Rephrasy’s documented HTML and Markdown support is worth a look. See Rephrasy’s tested configuration.

Rephrasy Superstealth passed 10 of 50 Pangram checks, zero Originality.ai checks and 22 GPTZero checks in the November 2026 HumanizerEval benchmark.

Rephrasy passed one in five Pangram tests

Pangram 4.1 gave a qualifying Human classification to 10 of Rephrasy’s 50 outputs. StealthGPT Super produced 34 more qualifying results on the same drafts. Rephrasy fared better on GPTZero at 22 passes and Winston AI at 44. Its weakest result came from Originality.ai 5.0, where none of the 50 outputs qualified as human. Rephrasy benchmark results.

AI detectorRephrasy Superstealth passesRephrasy Superstealth failuresStealthGPT Super passes
Pangram 4.110/50 (20%)40/50 (80%)44/50 (88%)
GPTZero 5.0b22/50 (44%)28/50 (56%)30/50 (60%)
Originality.ai 5.00/50 (0%)50/50 (100%)19/50 (38%)
Winston AI 5.244/50 (88%)6/50 (12%)48/50 (96%)
ZeroGPT20/50 (40%)30/50 (60%)34/50 (68%)
Weighted detector pass rate27.60%—69.00%
Rephrasy Superstealth and StealthGPT Super pass counts across all five detectors on the same 50 November 2026 benchmark drafts.

Why Rephrasy can pass Winston AI and still struggle on Pangram

Winston AI gave 44 Rephrasy rewrites a document score of 80 or higher on its Human scale. Pangram gave 10 rewrites its qualifying Human classification. The two columns count outcomes against different rules, which is why it helps to look at the specific detector a workflow uses. A row that says 0/50 counts how many documents qualified. Each detector can report different confidence scores or text classifications on those same documents. See the detector rules.

The benchmark weights Pangram at 40%, GPTZero at 30%, Originality.ai at 20% and Winston AI and ZeroGPT at 5% each. With those weights, Rephrasy scored 27.60% and StealthGPT Super scored 69.00%. That 41.4-point gap reflects the five-detector panel and the published weighting. See StealthGPT Super’s detector results.

Rephrasy scored 60.56 out of 100 for writing quality

Rephrasy’s strongest writing scores were completeness (93.60) and grammar (90.00). It came within 0.4 points of StealthGPT Super on completeness. The largest differences appeared in the benchmark’s checks for unsupported additions, structure and readability. StealthGPT Super received higher automated scores across all eight measures and finished third for writing quality. Rephrasy finished eleventh. See the writing quality leaderboard.

Writing quality measureRephrasy SuperstealthStealthGPT Super
Readability52.4068.80
Meaning preservation58.4067.20
Structure43.6062.00
Factual consistency64.0078.00
Completeness93.6094.00
Grammar and syntax90.00100.00
Control of unsupported additions2.0032.00
Stance preservation87.6094.80
Weighted writing quality60.5673.48
Rephrasy Superstealth and StealthGPT Super mean scores across all eight automated HumanizerEval writing checks, including the weighted total.

The automated fidelity check flagged 49 Rephrasy outputs for evidence review. A flag is a prompt for a human to compare the two drafts and trace additions back to the source. If your draft contains figures, named entities or a qualified claim, compare those first. A sentence can read smoothly even when a change to one number or qualifier alters the facts. Read how HumanizerEval scores quality.

Where Rephrasy’s HTML and Markdown support is useful

Rephrasy’s API documents HTML and Markdown modes for workflows that already have formatted drafts. The vendor’s example rewrites a heading and paragraph while retaining the HTML tags and code. The same documentation lists link destinations among the elements it preserves. That is useful when a draft starts as a CMS entry and needs to return as markup. Check heading text, anchor wording and facts as part of the review. The November results above apply to the tested Superstealth configuration. Read Rephrasy’s API documentation.

Is StealthGPT a better Rephrasy alternative?

StealthGPT Super was the stronger choice in this benchmark for both detector performance and writing quality. It earned more qualifying verdicts on all five detectors, scored 12.92 points higher on automated writing quality and finished two places higher overall. In particular, it recorded 44 Pangram passes to Rephrasy’s 10. For the full breakdown on the same 50 drafts, see StealthGPT vs Rephrasy.

If you’re working with an article, pulling a meaningful excerpt and checking the headline, citations, figures and qualifiers after rewriting is a practical first test. For an API publishing workflow, StealthGPT offers text preservation for spans you explicitly protect, such as citations, reference links and other exact strings. StealthGPT has published a separate six-call preservation test for headings, lists, tables, links and citations. Read the document preservation test.

How we compared Rephrasy and StealthGPT

HumanizerEval ran 15 services on the same 50 English inputs for its November 17, 2026 benchmark. It scored the 750 delivered outputs against five detectors and eight automated writing-quality checks. The overall score uses 70% weighted detector results and 30% writing quality. Mixed, AI and uncertain verdicts fail the strict pass rule. Winston AI uses a Human Score threshold of 80. The methodology documents post-run scoring changes, and the automated quality results are provisional. Read the full methodology.

FAQ

Does Rephrasy work?

In our November 2026 test, Rephrasy Superstealth finished fourth of 15 overall. It did well on Winston AI (44/50) and posted a 93.60 completeness score. It passed 10 Pangram checks, zero Originality.ai checks and scored 60.56 out of 100 on automated writing quality.

Does Rephrasy bypass Pangram?

Rephrasy did so on 10 of 50 drafts in the November benchmark. Pangram 4.1 gave StealthGPT Super a qualifying Human classification on 44 of those same 50 drafts.

Does Rephrasy bypass GPTZero?

Rephrasy Superstealth passed 22 of 50 GPTZero 5.0b checks, a 44% rate. StealthGPT Super passed 30, a 60% rate.

Does Rephrasy bypass Originality.ai?

None of Rephrasy Superstealth’s 50 rewrites received a qualifying human label from Originality.ai 5.0 in this test. StealthGPT Super passed 19 of the same 50 checks.

Does Rephrasy bypass ZeroGPT?

Rephrasy Superstealth passed 20 of 50 ZeroGPT checks. StealthGPT Super passed 34 of 50. GPTZero is a separate detector with its own results.

Can Rephrasy rewrite HTML and Markdown?

Yes. Rephrasy’s API documents HTML and Markdown input modes that preserve markup such as tags and link destinations while rewriting prose. See its structured input documentation.

What is a good alternative to Rephrasy?

StealthGPT Super is the stronger alternative for the tested mix of detector results and writing quality. It scored 70.34 overall to Rephrasy’s 37.49 and 73.48 on writing quality to Rephrasy’s 60.56. See the full head-to-head comparison.

Jozef Gherman
About the author
Jozef Gherman
CEO
Jozef is the founder and CEO of StealthGPT. Over three years ago he created StealthGPT with almost no money and in that short time has grown his business into a multimillion dollar phenomenon talked about and used globally.

StealthGPT, Your AI Humanizer for the Next Draft

Humanize essays, papers, and professional writing with StealthGPT. Explore our AI humanizer and compare the published humanizer test results.