Pangram AI text detector update claims one false flag per 24,000 documents

Pangram has launched Pangram 4, the newest version of its Pangram AI text detector, with API pricing set at $0.05 per 100 words and company-reported benchmarks claiming one false AI flag in about every 24,000 human-written documents. The product update is aimed at customers that need to screen text at scale, where false positives can be more damaging than missed detections, but the performance figures are Pangram’s own claims.

According to Pangram, Pangram 4 correctly identifies 99.66% of AI-generated text. The company also says the model incorrectly labels human-written text as AI-generated in 0.0041% of cases, which it frames as roughly one error per 24,000 documents.

The new model is six times larger than Pangram 3, Pangram said. Compared with the prior model, Pangram claims Pangram 4 produces 14 times fewer false positives and misses AI-written text six times less often.

What is Pangram 4?

Pangram 4 is a model for detecting whether text was written by a human, generated by AI, or affected by AI editing. Pangram says the model can separate lightly AI-polished writing from text that was generated fully by AI, a distinction that matters for customers trying to enforce policies without treating every use of an assistant as the same event.

The company also says Pangram 4 is resistant to tools that rewrite AI output to make it appear more human. Pangram tested the model against 13 common “humanizers” and reported that it still detected the AI component in 98.83% of cases.

What changes for Pangram customers?

The API now costs $0.05 per 100 words. Pangram said that means a price increase of two to ten times, depending on document length. The company did not describe a single universal price change, which makes the impact dependent on how customers submit text and how long those documents are.

All plans can now scan images at no extra charge, according to Pangram. That addition expands the scope of what customers can submit through the product, though Pangram’s announcement centers on text-detection performance rather than image-specific benchmark figures.

Pangram 3 is not being discontinued immediately. Pangram said it will keep supporting the prior model for two more months, with shutdown scheduled for Sept. 30, 2026.

The update gives Pangram a cleaner enterprise sales argument: better accuracy, fewer false accusations, and better handling of AI-edited or AI-obfuscated text. The tradeoff is price. For teams using detection across large document volumes, the cost increase may matter as much as the reported accuracy gains.

This story draws on original reporting from The Decoder.