Pangram Labs launched Pangram 4, a new AI-text-detection model the company says is roughly six times the size of its predecessor and flags AI-assisted writing with a false positive rate of 0.0041 percent, or about one wrong flag per 24,000 documents. False accusations of AI authorship carry real costs: students face academic-integrity hearings and job seekers lose offers over a detector score, so a vendor’s own accuracy claim deserves scrutiny rather than acceptance on the strength of a blog post.
Pangram Labs attributes the 0.0041 percent figure to its own internal benchmarks, not an outside audit. The company also reports a false negative rate of 0.3396 percent, down from 1.99 percent for Pangram 3, measured on what it calls new challenge datasets. Neither figure comes with a stated base rate of AI-written text in the test population or a description of how the documents were collected, and a false-positive rate only means something against a defined test set and base rate.
Pangram Labs does offer more transparency than many detector vendors: a technical report, a model card, and a separate technical blog post accompany the launch, and the company points readers toward that methodology rather than the headline number alone. That is the right instinct for a category where wrongful flags carry real consequences, even though no outside researcher appears to have replicated the results yet.
The new model targets a specific failure mode: separating AI-edited human writing from documents that interleave human and AI passages, something Pangram Labs says no earlier detector could handle in a single pass. The company reports catching AI involvement in humanized writing 98.83 percent of the time when tested against 13 commercial humanizing tools, and it says every frontier model family it tested was caught with a false negative rate below 0.7 percent, with detectability showing no tie to when a given model was released.
Pangram Labs is also overhauling pricing alongside the launch. Credits now draw down per 100 words scanned rather than rounding up to 1,000, image scanning joins every plan, and API pricing moves to $0.05 per 100 words, up to tenfold higher than before for long documents. Pangram 3 keeps running until its September 30, 2026 deprecation.
Any school, employer, or platform evaluating this detector should ask Pangram Labs for the underlying test set and base rate behind the 0.0041 percent claim before treating it as protection against wrongful accusations, and should model the higher API costs on long documents before committing to a 2026 contract.
Pangram Labs disclosed these figures in its own blog post introducing Pangram 4, published on pangram.com without a dateline and reviewed July 30, 2026.
