Introducing Phrasly UltraWatermark-free. Industry-leading pass rates.Explore Ultra
Release Notes

Changelog

Every meaningful update to Phrasly's AI models and writing tools, dated and in one place: new model versions, quality and readability improvements, language support, and fixes. All models are built and trained in-house, and performance numbers cited here come from our internal evaluations at the time of each release.

Last updated:

2026
Added
phrasly-ultra

Phrasly Ultra

A new humanizer model, trained and served in-house, available to paid plans and trials.

  • Model quality
    • Averaged a 99.8% Pangram human score across 72 AI-generated English texts in our September 2026 benchmark, with 70 of 72 outputs scoring 100% human and no refusals.
    • Inputs were sampled from the public AI Text Detection Pile dataset and scored with Pangram (pangram-4). Results vary by text.
  • Rewrite modes
    • Five rewrite approaches, from Gentle to Max, selectable on a slider in the humanizer. Balanced is the default and the recommended starting point.
    • Your choice is remembered between sessions in your browser.
  • Where it runs
    • Available in the humanizer, the document editor, the content generator and the Phrasly MCP server. Tools without a model picker use the Balanced mode.
    • Works in any language Phrasly supports: Ultra rewrites text in the language you wrote it in, rather than routing non-English text to a separate model. Multilingual coverage is not yet tested across every language, so results vary by language and text.
    • Included with every paid plan and with the trial. Free accounts continue on the Balanced model.
Removed

Legacy humanizer model retired

  • The legacy humanizer model has been switched off and no longer serves requests. It had been unavailable in the model picker since the picker was introduced.
  • Accounts still carrying a saved Legacy preference are moved to the Balanced model automatically. No action is needed.
Changed

Word limit doubled to 5,000 words

  • Raised the per-run word limit from 2,500 to 5,000 words for all users.
Added

MultiLang Humanizer model

A new multilingual humanizer model, trained and evaluated in-house.

  • Supports 25 languages, including Spanish, French, German, Portuguese, Chinese, Hindi, Arabic, Japanese, Korean, and Russian.
  • Grammar quality improved by more than 200% over the previous multilingual model on our internal evaluation set.
  • Higher human-likeness scores across all supported languages in internal detector evaluations (January 2026).
Changed

Output consistency improvements

  • Improved the consistency and reliability of model outputs.
Added

New humanizer model and model picker

  • Model quality
    • 59.1% improvement in detection scores compared to the previous model, measured on our internal benchmark across the major AI checkers (January 2026).
    • Resolved a regression against GPTZero’s updated paraphrasing detection that affected some texts.
  • Model picker
    • You can now choose between the latest model and previous versions. Earlier models can read as more formal, which some users prefer for professional writing.

These features are live as of 5 PM EST. Contact [email protected] if you run into any issues. More updates are planned for later this month. - Daniel

2025
Info

Turnitin product update: compatibility review

  • Following Turnitin’s latest product update, we re-evaluated the current humanizer model. No model changes were required. (Turnitin update notes)
  • Regression test: 100 documents processed through the humanizer were checked against Turnitin in August 2025; all 100 returned a 0% AI score.
  • Because Phrasly runs its own models rather than reselling a third-party AI service, watermarking changes made by outside providers do not carry into Phrasly output.
Added
v6-2025

V6 Humanizer

A full model refresh focused on speed and naturalness.

  • Processes text up to 3x faster than previous versions.
  • Generates more natural text that adapts better to the tone of the original.
  • Improved detection scores across the board in our June 2025 internal benchmark.
  • Resolved detector-specific regressions reported by users:
    • GPTZero’s updated paraphrasing detection, which affected certain texts.
    • GPTZero’s AI-editing detection.
    • Improved results on Copyleaks.

We’re committed to keeping our humanizer model the best in the industry. We look forward to your feedback!

Changed
v5.1-2025

Readability tuning

  • Improved readability, with small adjustments to detection performance.
Changed
v5-2025

V5 model

A new model release focused on output quality.

  • Adapted the model to GPTZero’s paraphrasing-detection update.
  • Improved detection scores across the other major AI checkers in internal testing.
  • More natural sentence flow and phrasing.
  • Fixed grammar errors that could appear in output.
2024
Changed
2024-12-01-v3

Aggressive and Medium modes reworked

The two humanizer modes now make their trade-off explicit.

  • Aggressive: prioritizes detection performance first, which can reduce readability.
  • Medium: prioritizes readable drafts while keeping strong detection performance in our testing.
Added
2024-11-03-v2

New AI Humanizer architecture

  • Added sentence variations, giving you more freedom to edit your humanized text.
  • Fully revamped the humanizer model for higher-quality output.
  • Fixed reports of humanizer output running significantly longer than the input. Output length now stays nearly identical to the original.
Added
2024-01-17-base

Fully in-house humanizer model

  • Moved the humanizer onto models we develop and train entirely in-house. Changes made by outside AI providers, such as OpenAI’s announced watermarking plans, do not affect Phrasly output. (about watermarking)
  • Benchmarked the new model against the major AI checkers, including Turnitin, before release.
2023
Fixed
v1.2

Long-text quality fix

  • Improved text quality for longer inputs. Quality no longer degrades as the input gets larger.
Added
v1

Initial public release

  • Released the AI Humanizer model to the public.

Have an issue or some feedback? Contact us at [email protected].