Phrasly Ultra Beats GPTZero 5.0b: 97% Pass Rate and Our Best Writing Quality Yet
GPTZero's November 4 update was built to catch humanized text. Eight days later the updated Phrasly Ultra passes it 97% of the time, up from 63%, with a grammar score of 9.0 out of 10. Here are the before and after numbers.
4 min read
Phrasly Ultra launched on September 15 as our flagship humanizer and went straight to the top of our benchmarks. On November 12, 2026, we replaced it with something better. The updated Ultra is live now for every paid plan and trial, everywhere Ultra runs. There is nothing to switch on.
The update had two goals: beat GPTZero's new detection model, which started catching humanized text on November 4, and make the writing itself cleaner. It does both, by a wide margin. This is the strongest humanizer we have ever released, and the numbers below show how far it moved, measured on the same texts before and after.
The update at a glance
Measured on the same texts | Previous Ultra | Updated Ultra |
|---|---|---|
GPTZero 5.0b pass rate | 63.4% | 97.1% |
GPTZero 5.0b pass rate, long documents | 38.9% | 97.2% |
Pangram pass rate | 96.5% | 95.8% |
Grammar, out of 10 | 5.43 | 9.05 |
Fluency, out of 10 | 6.42 | 8.98 |
Meaning kept, out of 10 | 6.70 | 7.43 |
Rewrites with a serious error | 14.4% | 3.7% |
Every row compares the two models on identical texts, in Balanced mode, with one attempt per text. No retries and no picking the best draft.
GPTZero 5.0b caught humanizers. Ultra passes it.
When Ultra launched in September, it passed GPTZero more than 99% of the time. Then, on November 4, GPTZero released Model 5.0b, an update built to catch polished and paraphrased AI text. It worked. Text from humanizers that had passed the week before started coming back flagged, and Ultra was no exception: on 5.0b, the previous Ultra's pass rate fell to 63.4%.
Eight days later, the updated Ultra passes GPTZero 5.0b 97.1% of the time across 920 real-world texts: essays, reports, emails, articles and the rest of what people bring to a humanizer. To our knowledge, that makes Phrasly Ultra the first humanizer to pass GPTZero's new model.
We count a pass the hard way. GPTZero has to label the whole text human. A "mixed" verdict counts as a fail.
Pangram, where Ultra already led, stays where it was: 95.8% across 1,920 texts against 96.5% before, a gap inside the margin of error of the test. You are not trading one detector for another. You keep Pangram and get GPTZero back.
Writing you do not have to fix
Passing a detector is half the job. The other half is getting back text you can use as it is. This is where the update moved the most.
Grammar: 9.05 out of 10, up from 5.43. Commas, articles, agreement and sentence boundaries come back the way a careful writer would leave them.
Fluency: 8.98, up from 6.42. Rewrites read naturally from the first sentence to the last.
Meaning: 7.43, up from 6.70. More of what you wrote survives the rewrite, including names and figures.
Serious errors: 3.7%, down from 14.4%. A serious error is a rewrite that lost part of the meaning, dropped a number or repeated itself. That was about one rewrite in seven. It is now about one in twenty-seven.
Punctuation that matches your original. The updated Ultra's rewrites carry about 78% as many commas as the text you put in. The previous model's carried about 13%.
In practice that means fewer rereads, fewer manual fixes and fewer runs spent regenerating a draft that was almost right.
Long documents: the biggest jump
Long texts are the hardest case for any humanizer. There is more room to drift and more for a detector to work with. On 72 documents of 1,800 to 5,000 words, the update changes the picture completely.
72 long documents | Previous Ultra | Updated Ultra |
|---|---|---|
GPTZero 5.0b pass rate | 38.9% | 97.2% |
Pangram pass rate | 90.3% | 97.2% |
Rewrites with a serious error | 47.2% | 5.6% |
If you humanize full essays, reports or chapters in one run, this is the part of the update you will feel first.
What changes for you
Nothing to turn on. Every Ultra humanization has used the updated model since November 12: in the humanizer, in Neo, in the document editor and through the Phrasly MCP server for Claude and ChatGPT.
The same five modes. Gentle, Light, Balanced, Strong and Max work as before. Balanced is still the default and still the place to start.
One setting retired. "Increase comma count" is gone, because the updated Ultra keeps your commas without it. Sentence length, Cleaner grammar and Proofread are unchanged.
Same plans, same limits, same price. Ultra is included with every paid plan and with the trial.
How we tested
All of these numbers come from an internal evaluation we ran in November 2026.
The texts. 1,920 real-world texts of up to 5,000 words, including 72 long documents. All 1,920 were scored by Pangram, 920 of them by GPTZero as well, and 1,000 were graded for writing quality.
The runs. Both models rewrote every text once, in Balanced mode with default settings.
Detection. Every GPTZero score is from Model 5.0b, the version released on November 4. A rewrite passes only when the detector's verdict for the whole text is human. A mixed verdict is a fail.
Quality. An automated grader scored each rewrite for grammar, fluency and meaning out of 10 and flagged serious errors. The same grader and the same instructions were used for both models.
Results vary by text, and no humanizer can promise a specific result on a specific detector. What we can promise is that we measure, we publish what we find, and we keep improving the model.
Try the updated Ultra
The updated Phrasly Ultra is live now. Paste a draft, leave it on Balanced, and see the difference in the first rewrite.
Frequently asked questions
Phrasly Ultra was replaced with an updated model on November 12, 2026, built to pass GPTZero's new Model 5.0b. In our internal evaluation its GPTZero pass rate rose from 63.4% to 97.1%, its grammar score rose from 5.43 to 9.05 out of 10, and rewrites with a serious error fell from 14.4% to 3.7%, while its Pangram pass rate stayed level.
Written by
