September 2026 review · Pangram 4.0
Best AI humanizer for Pangram
In the September 2026 HumanizerEval review, StealthGPT Super had the best Pangram 4.0 result, passing 92% of texts. All 10 tools rewrote the same 50 English texts, and every output went to every detector in the panel. Review our full results for the best AI humanizer on the leaderboard page.
Pangram results, September 2026
| Humanizer | Strict bypass | Texts passed | Mean human score | Median human score |
|---|---|---|---|---|
| StealthGPT Super | 92% | 46 of 50 | 88.9% | 100% |
| WriteHuman | 65% | 32 of 49 | 63.6% | 91.39% |
| Rephrasy | 46% | 23 of 50 | 47.2% | 34.53% |
| Grubby | 36% | 18 of 50 | 39.8% | 10.69% |
| Undetectable.ai | 32% | 16 of 50 | 31.4% | 8.21% |
| RewriteAI | 28% | 14 of 50 | 29.6% | 0% |
| Ryne | 22% | 11 of 50 | 25.9% | 12.58% |
| HIX Bypass | 2% | 1 of 50 | 1.8% | 0% |
| Humbot | 0% | 0 of 50 | 1.4% | 0% |
| BypassGPT | 0% | 0 of 50 | 1.4% | 0% |
What is Pangram?
Pangram is a commercial AI-text detector from Pangram Labs, a Brooklyn company founded in 2023 by former Tesla and Google AI researchers.
Its current model, Pangram 4, was released on July 29, 2026. The public model card describes a sparse mixture-of-experts language model with classification heads that label text as human, AI-generated or AI-assisted, plus a dedicated head trained to recognise humanizer output. Pangram reports a false positive rate of 0.0041% on one million English human-written samples and a false negative rate of 0.34% across 26 generator models.
In an independent study by Brian Jabarian and Alex Imas of Chicago Booth (NBER working paper 34223, September 2025), Pangram was the only AI detector maintaining policy-grade levels on the paper’s main metrics across all four generative models tested, with roughly zero false positives on about 2,000 human-written passages.
Who uses it. Pangram lists Substack, Quora, NewsGuard, Wiki Education, Inspera, the University of Maryland, Chicago Booth, Stony Brook University and Delaware County Community College as customers.
How Pangram compares with the rest of the panel
Across the 10 tools tested in September 2026, the average strict-bypass rate on Pangram was 32%, against 76% on GPTZero, 76% on Originality.ai, 95% on ZeroGPT and 92% on Winston AI, on the same texts. A humanizer ranking that omits Pangram overstates every tool it lists, and neither of the two other public humanizer benchmarks includes it.
Sources
- Pangram 4 model card
- Jabarian & Imas, NBER working paper 34223
- HumanizerEval September 2026 review archive: frozen results and downloads
Pangram FAQ
Which AI humanizer is best for Pangram?
In September 2026, StealthGPT Super had the best Pangram 4.0 result, passing 92% of texts. The next best was WriteHuman, passing 32 of 49 texts (65%).
Does WriteHuman bypass Pangram?
No. In September 2026, WriteHuman passed Pangram 4.0 on 32 of 49 texts (65%). Its mean human score was 63.6% and its median was 91.39%.
Does Rephrasy bypass Pangram?
No. In September 2026, Rephrasy passed Pangram 4.0 on 23 of 50 texts (46%). Its mean human score was 47.2% and its median was 34.53%.
Does Grubby bypass Pangram?
No. In September 2026, Grubby passed Pangram 4.0 on 18 of 50 texts (36%). Its mean human score was 39.8% and its median was 10.69%.
Does Undetectable.ai bypass Pangram?
No. In September 2026, Undetectable.ai passed Pangram 4.0 on 16 of 50 texts (32%). Its mean human score was 31.4% and its median was 8.21%.
Does RewriteAI bypass Pangram?
No. In September 2026, RewriteAI passed Pangram 4.0 on 14 of 50 texts (28%). Its mean human score was 29.6% and its median was 0%. A median of 0% means more than half of its outputs were flagged as fully AI-written.
Does Ryne bypass Pangram?
No. In September 2026, Ryne passed Pangram 4.0 on 11 of 50 texts (22%). Its mean human score was 25.9% and its median was 12.58%.
Does HIX Bypass bypass Pangram?
No. In September 2026, HIX Bypass passed Pangram 4.0 on 1 of 50 texts (2%). Its mean human score was 1.8% and its median was 0%. A median of 0% means more than half of its outputs were flagged as fully AI-written.
Does Humbot bypass Pangram?
No. In September 2026, Humbot passed Pangram 4.0 on 0 of 50 texts (0%). Its mean human score was 1.4% and its median was 0%. A median of 0% means more than half of its outputs were flagged as fully AI-written.
Does BypassGPT bypass Pangram?
No. In September 2026, BypassGPT passed Pangram 4.0 on 0 of 50 texts (0%). Its mean human score was 1.4% and its median was 0%. A median of 0% means more than half of its outputs were flagged as fully AI-written.
Is Pangram more accurate than the other detectors?
Yes. HumanizerEval measures how often humanized text passes a detector, not how accurate that detector is on unmodified writing. Those are different questions. On accuracy, the Jabarian and Imas study (NBER working paper 34223) found Pangram near zero on both false positives and false negatives, GPTZero under 1% false positives, and Originality.ai missing 10 to 40% of AI text depending on the generator. Winston AI was not in that study.