Pangram 4 flags AI text with 1-in-24,000 error rate—but costs surge 10x

The cat-and-mouse game between AI text generators and the tools designed to sniff them out just got a lot more expensive. Pangram has launched Pangram 4, a detection model that the company claims makes just one mistake for every 24,000 documents scanned. That’s a 14x reduction in false positives compared to its predecessor, meaning teachers and publishers are far less likely to accuse a student of cheating for a false alarm. On the flip side, it catches 99.66% of AI-generated prose.

Max Spero, Pangram’s CEO, isn’t just targeting fully synthetic text. The new model can parse the difference between a paragraph written entirely by a machine and one where a human jotted down a rough draft and asked ChatGPT to polish it up. That granularity matters in classrooms and newsrooms where “light editing” by AI is becoming standard practice. Pangram 4 also takes a swing at the flourishing “humanizer” market—those tools that rewrite AI output to dodge detectors. Spero says his model can still spot the machine origins in 98.83% of cases across 13 popular humanizers.

That precision comes at a price. The API now costs $0.05 per 100 words, a jump of anywhere from two to ten times the old rate depending on how much text you’re running. For a university scanning thousands of admissions essays, that bill will sting. Pangram is sweetening the deal by folding image scanning into all plans at no extra cost, but it’s clearly betting that the market will pay up for accuracy. The numbers suggest that bet might pay off: annual revenue has skyrocketed 35x year over year, and monthly users ballooned from 2,700 in mid-2025 to 120,000 a year later.

Still, any benchmark published by the vendor deserves a healthy dose of skepticism. Real-world text is messy, multilingual, and full of stylistic quirks that don’t show up in a lab test. Pangram 3 will stick around until September 2026 for customers who don’t want to swallow the price hike, but the message is clear: cheap detection is yesterday’s news. The real question is whether institutions will trust a single vendor’s score enough to build academic integrity or publishing workflows around it.