StealthGPT Super finished first in the September 2026 HumanizerEval review, with a 96.8% mean strict bypass and 46 of 50 texts clearing all five detectors. That’s a strong headline. A careful reader still wants the details behind it, including where the tool is weaker.

Here’s what the profile page shows, how the vendor’s own claim held up, and what to keep in mind when you read a result from a tool tied to the people running the test.

What Did StealthGPT Super Score on Each Detector?

On Pangram, it passed 92% of texts, with an 88.9% mean human score and a 100% median. On GPTZero it passed 98% with a 99.0% mean, on Originality.ai 98% with a 95.8% mean, on ZeroGPT 98% with a 91.1% mean, and on Winston AI 98% with a 97.7% mean.

Those are consistent numbers. Many tools do well on lenient detectors and fall off on Pangram, but this one stayed within six points across all five. The next best Pangram pass rate was WriteHuman’s 65%, so on the hardest column the gap to second place was 27 points, well beyond the margin of error.

How Did the Vendor Claim Hold Up?

The profile page quotes StealthGPT’s own statement, retrieved September 18, 2026, that it passes Pangram at an 89% bypass rate based on internal testing. The measured Pangram result was 92%, so the verdict reads Verified. It’s a useful format, since it sets marketing and measurement side by side.

Read it with proportion. At 50 texts, a 95% confidence interval is roughly plus or minus 14 points, so 89% and 92% are effectively the same figure. When you compare top AI humanizer services, a claim that matches the measurement is reassuring, though it doesn’t prove performance on texts the review never saw.

What Does the Quality Table Say?

The strengths

StealthGPT Super had the best mean quality pass rate in the panel at 83.5%. Syntax integrity was 98% and stance preservation 96%, so sentences stayed grammatical and the author’s position was rarely distorted. Length stayed close to the original, with a ratio of 1.04 and no output beyond 1.4 times or below 0.6 times the input.

The limits

Factual consistency was 68% and naturalness 72%. That means roughly a third of outputs had an issue with claims, numbers or names, and more than a quarter had awkward phrasing. The profile page answers “Does it change the meaning?” with a no, but its own figures show that meaning changes still happen in a meaningful share of outputs.

How Did It Do Head to Head?

The matchups are lopsided. StealthGPT cleared all five detectors on 14 samples where WriteHuman did not, and the reverse happened once, with 31 shared. Against Grubby, the counts were 29 and 1 with 17 shared, and against Rephrasy 26 and 0 with 20 shared. Against the bottom four, the leader won 45 or 46 samples with no reverse.

Per sample agreement is a sharper test than averages, since every tool saw the same inputs. Still, the tool is measured at one setting, Super, using production output, on one date. A different mode or a later detector update could change those counts, so treat them as a snapshot.

What to check before relying on this profile

• The settings rule used for the run

• The detector versions recorded

• Your own documents, tested side by side

• The quality columns, not just bypass

Why Does Provenance Matter Here?

The public data repo sits under the StealthGPT Labs account, and the tool profiled here is the one ranked first. That’s an overlap you should notice and name, especially since the same site criticizes two other benchmarks for ranking their own operators first. The overlap doesn’t make the numbers false.

What it does is raise the value of the open data. The CSV holds 50 rows of anonymized per sample scores, a script recomputes every aggregate, and the SHA-256 hash lets you confirm the file is unchanged. Use those tools before you let any single profile steer a purchase.

How Should You Use This Profile in Practice?

Use it to set expectations. If your audience runs Pangram, the 92% pass rate is a strong sign. If your content carries figures and citations, the 68% factual consistency says you’ll need a careful read after every rewrite. And if your brief has a tight word limit, the 1.04 length ratio is a real advantage.

Then run your own test. This AI humanizer comparison gives you an informed starting point, but five or ten of your own documents, tested with identical settings, will tell you more about fit. Keep a log, mark every changed fact and revisit the result when detectors update.

Conclusion

StealthGPT Super earned its first place on the September 2026 data, with strong bypass across all five detectors and the best quality pass rate. Its factual consistency and naturalness numbers keep the story honest. Verify the data, mind the provenance and test on your own text.

FAQs

What was StealthGPT Super’s Pangram pass rate?
92%, or 46 of 50 texts, with a mean human score of 88.9%.

Did the vendor claim match the measurement?
Yes. It claimed 89% on Pangram, and the review measured 92%, labelled Verified.

What was its factual consistency score?
68%, so about a third of outputs had an issue with claims, numbers or names.

Did it make text much longer?
No. Its mean length ratio was 1.04.