November 2026 benchmark
How our models score on AI detectors.
Average human scores on 60 fresh AI texts, one try each.
| Detector | Phrasly Ultra | Phrasly Lite |
|---|---|---|
| Pangram | 99.2% | 95.1% |
| GPTZero | 94.7% | 87.6% |
| Originality.ai | 84.2% | 38.4% |
| ZeroGPT | 100% | 90.3% |
| Average | 94.5% | 77.9% |
876 real-world texts
The scores at the top of this page come from 60 texts written for the test. These 876 were not: they are real-world texts set aside for testing, among them 48 documents of 1,000 to 1,800 words and 28 texts in other languages. Nobody picked them for being flagged, so the last column shows how they scored before humanizing.
| Detector | Phrasly Ultra | Phrasly Lite | Original text |
|---|---|---|---|
| Pangram | 99.5% | 95.2% | 18.2% |
| GPTZero | 98.7% | 95.9% | 16.4% |
| Originality.ai | 88.0% | 60.7% | 17.7% |
| ZeroGPT | 98.5% | 86.9% | 13.3% |
| Average | 96.2% | 84.7% | 16.4% |
Phrasly Ultra ran in Balanced mode. Originality.ai checks texts of 30 words or more, so its averages leave out 8 of Ultra's rewrites and 6 of Lite's. The detector checks ran from November 12, 2026 to November 17, 2026.
Writing quality
A rewrite that gets past a detector still has to read well and say what the original said. On the same 876 texts, an automated grader scored every rewrite out of 10, and a style check counted odd shortcuts: slashes between words, ampersands, clipped words and bracketed asides.
| Measure | Phrasly Ultra | Phrasly Lite |
|---|---|---|
| Grammarout of 10 | 9.20 | 9.34 |
| Fluencyout of 10 | 9.01 | 9.09 |
| Meaning keptout of 10 | 7.25 | 7.19 |
| Odd shortcutsper 100 words, lower is better | 0.1 | 0.3 |
By source chatbot
Phrasly Ultra's scores on the 60 fresh texts, split by the chatbot that wrote the original, 10 texts each. The lowest is 64.6%, from Originality.ai on texts by GPT-6.1 Terra. Across the four detectors, no chatbot's texts average below 91.1%.
| Original written by | Pangram | GPTZero | Originality.ai | ZeroGPT | Average |
|---|---|---|---|---|---|
| GPT-6.2 Sol | 100% | 100% | 86.8% | 100% | 96.7% |
| GPT-6.1 Terra | 100% | 100% | 64.6% | 100% | 91.1% |
| Claude Sonnet 5.5 | 100% | 84.9% | 86.6% | 100% | 92.9% |
| Gemini 4 Argon | 95.4% | 89.0% | 86.2% | 100% | 92.6% |
| Grok 4.7 | 100% | 100% | 94.4% | 100% | 98.6% |
| DeepSeek V4.1 Flash | 100% | 94.3% | 86.6% | 100% | 95.2% |
Text by text
One square per text, in the same order for both models. The darker the square, the more of the four detectors returned a human verdict for that rewrite. Hover or tap a square to see each verdict.
Method
- Models
- Phrasly Ultra, our flagship AI humanizer, in Balanced mode with every text setting at its default. Phrasly Lite, our free AI humanizer, which has one mode. Ultra also ran in Strong and Max; this page reports Balanced, its default.
- Texts
- 60 AI texts written for this test by 6 chatbots, 10 each: GPT-6.2 Sol, GPT-6.1 Terra, Claude Sonnet 5.5, Gemini 4 Argon, Grok 4.7, and DeepSeek V4.1 Flash. They cover 10 kinds of writing, from essays to cover letters, at 147 to 1,171 words. Every detector flagged all 60 before humanizing; their average human score was 0.9%.
- Attempts
- One rewrite per text. No retries, and no choosing between drafts.
- Detectors
- Pangram, GPTZero, Originality.ai and ZeroGPT, each called through its API on November 17, 2026. Model versions the APIs reported: Pangram 4.1 and Originality.ai 5.0.
- Score
- The average human score: each detector's own human score for the whole text, averaged over the texts. A mixed verdict counts for what it scores, not as a fail. The figure for all four is the average of the four detectors' averages.
Each detector's human score
| Detector | Human score used |
|---|---|
| Pangram | The share of the text it reads as human. Passages it reads as AI-assisted count against the score. |
| GPTZero | Its probability that the whole text was written by a human. |
| Originality.ai | 100 minus its AI score for the text. |
| ZeroGPT | 100 minus the percentage of the text it reads as AI. |
Our technical write-up of this benchmark goes further: how the two models work, how precise these averages are, and what a rewrite does to length and meaning.
Limits
- No Turnitin
- Turnitin has no public API, so we could not run it under these rules.
- Originality.ai is the hardest
- It is the lowest score for both models on both sets. Lite's falls with length: 57.3% on the shortest third of the 60 fresh texts, 20.3% on the longest. Ultra's holds better, at 86.6% and 82.6%.
- Small samples
- 60 texts make the averages approximate: the 95% interval is 91.7% to 96.9% for Ultra's 94.5%, and 73.5% to 82.0% for Lite's 77.9%. Each chatbot's row rests on 10 texts, so read those as rough.
- A dated snapshot
- Detectors update their models, and scores move when they do. These are the scores of November 17, 2026.
- Averages, not promises
- Scores vary from text to text, as the squares above show. Run your own rewrite through our AI detector before you rely on it.
- Our models only
- This edition has no other humanizers. The September 2026 edition, under Data, compared Ultra with StealthGPT, WriteHuman and Undetectable AI.
Guarantee and $500 challenge
$500 Beat Ultra challenge
Show another humanizer beating Phrasly Ultra on average across at least 10 random real-world AI texts, and we send you a $500 gift card. One-off texts don't count.
Challenge rulesMoney-back guarantee
If another humanizer scores more human than Ultra on the same text of 500 words or more, we refund your latest payment.
Guarantee termsBoth are judged on Pangram, GPTZero, Copyleaks and Originality.ai, and the other tool has to lead on at least two of them.
Data
Each file lists the 60 input texts with the chatbot and prompt that wrote them, and every detector's verdict and human score before and after humanizing.
September 2026 edition
Phrasly Ultra against StealthGPT, WriteHuman and Undetectable AI on 72 texts, scored by Pangram on September 15, 2026.
Questions
The one that scores highest on the detectors your reader uses, on text like yours. In this benchmark, Phrasly Ultra averaged 94.5% human across four detectors on 60 fresh AI texts. If you find a humanizer that beats it, the $500 challenge pays you to show us.