Meet model 4: fewer false flags on short, edited text
Our newest detection model is live for every account. It's better on short passages, lightly edited AI drafts and writing from non-native English speakers.
Jonas Weber
Product manager · Sep 18, 2026 · 4 min read
Model 4 is rolling out to every Winston AI account this week, on every plan and through the API. You don't need to change anything: scans started after the update use the new model automatically.
What changed
- Shorter text: more reliable results on passages between 150 and 400 words.
- Edited drafts: better at spotting AI text that has been lightly reworded by hand or by a paraphrasing tool.
- Fairer to second-language writers: a much larger training set of human writing from non-native English speakers.
- Newer models: training data now includes output from the latest releases of the major model families.
What stayed the same
Reports look the same, scores mean the same thing, and your history is untouched. Earlier scans keep the score they were given. If you want a fresh read, open the document and scan it again.
The goal for this release was simple: just as good at catching AI, and noticeably more careful with people.
Read the numbers
We've updated our public benchmark with model 4's results, including recall per model family and false positive rates broken down by text length. The full methodology is on our research page.
Written by
Jonas Weber, Product manager
Jonas leads the scan and reporting experience, from the first paste to the shared report.