Today we’re releasing Proof5 2, the second generation of our AI text detection model. It is four times larger than the model it replaces and was trained on a much wider set of writing, including essays, news, reviews, fiction and text that has been run through paraphrasing tools.
This post covers what changed, how it performs, and what it means for your plan. For full evaluation details, see the technical report.
Results
On our internal benchmark of 40,000 documents, Proof5 2 wrongly flags roughly one human document in every 20,000. It also misses less AI text than before, particularly from the newest frontier models.
Placeholder figures for layout.Mixed authorship
Most real writing isn’t purely human or purely AI. Proof5 2 reads a document as a whole and marks where AI-written passages begin and end, so a reviewer can see that the introduction was drafted by hand while the conclusion was not.
A score alone starts an argument. Highlighted passages start a conversation.
Pricing changes
Checks are now counted per 100 words instead of rounding up to 1,000, so short documents use far fewer credits.
- Free: 5 checks a day, up to 1,000 words each
- Pro: up to 500,000 words a month
- API: $0.04 per 100 words
What’s next
Proof5 1 remains available until January 31, 2027. Requests that don’t name a model will move to Proof5 2 on that date.
Maya leads model training and evaluation at Proof5. Before that she worked on language models for search and translation.
More from Maya →