Pangram 4 claims one error per 24,000 documents in AI detection
Pangram has released Pangram 4, a new version of its AI text detector that the company says is six times larger than its predecessor, Pangram 3, with 14 times fewer false positives and six times fewer missed AI-generated texts. In benchmarks run by Pangram itself, the model correctly identifies 99.66 percent of AI-written text and falsely flags human writing only 0.0041 percent of the time, roughly one error in every 24,000 documents.
The new version also distinguishes text that has only been lightly AI-polished from text that is fully AI-generated, according to Pangram. It is built to resist "humanizer" tools, services designed to disguise AI writing as human-written; Pangram says it still catches the AI component across 13 common humanizers 98.83 percent of the time.
Pricing for the API rises to $0.05 per 100 words, a two- to tenfold increase over the previous rate depending on document length. All plans now include image scanning at no extra charge. The older Pangram 3 remains available to customers until September 30, 2026.
Citing the New York Times, Pangram describes its business as booming: annual revenue has grown 35 times year over year, and monthly active users rose from 2,700 in June 2025 to 120,000 in June 2026.
Key facts
- Pangram 4 is six times larger than predecessor Pangram 3, with 14 times fewer false positives and six times fewer missed AI texts.
- Self-reported benchmarks: 99.66 percent correct identification of AI text and a 0.0041 percent false-positive rate, about one error per 24,000 documents.
- Catches AI-authored content run through 13 common "humanizer" tools 98.83 percent of the time, and can separate lightly AI-polished text from fully AI-generated text.
- API price rises to $0.05 per 100 words, a two- to tenfold increase; all plans now include free image scanning; Pangram 3 stays available until September 30, 2026.
- Per the New York Times, Pangram's annual revenue is up 35x year over year and monthly users grew from 2,700 in June 2025 to 120,000 in June 2026.
Why it matters
AI text detectors live or die by their false-positive rate, since wrongly accusing a human writer of using AI is the complaint that has dogged the category. Pangram 4's claimed 14-fold cut in false positives, alongside fewer missed AI texts, targets exactly that weak spot at a moment when AI-generated writing keeps growing across schools, newsrooms and workplaces.
Who it affects
Educators and institutions checking student work, publishers and newsrooms screening submissions, and hiring teams reviewing applications all rely on tools like this, as does anyone whose own writing risks being wrongly flagged as AI-generated. Makers of "humanizer" tools, which try to disguise AI writing as human, are the other side of this contest.
How to use it
Pangram 4 is available through Pangram's API at $0.05 per 100 words, a two- to tenfold increase over the prior rate depending on document length. Image scanning is now bundled into all plans at no extra cost. Customers not ready to switch can keep using Pangram 3 until September 30, 2026.
How solid is it
The 99.66 percent and 0.0041 percent figures come from Pangram's own benchmarks, not an independent third-party audit, so they should be read as the company's claim about its own product rather than a verified external result. The business figures, 35x revenue growth and the jump in monthly users, are attributed to the New York Times, which gives that part of the story some outside framing even though the detection-accuracy numbers remain self-reported.
Risks and caveats
Self-reported benchmarks can look stronger on a company's own test set than they do across the wider mix of genres, languages and writing styles the tool is not tuned for. The price increase, up to tenfold over the previous tier, could strain schools, publishers or other high-volume customers, and raises the bar for smaller users who relied on cheaper detection.