Pangram Labs launches most accurate AI text detector with 100% adversarial detection

Pangram Labs, a team of over 25 full-time employees, claims to have built the most accurate AI text detector in the world. According to their source paper, their baseline model achieves 100% accuracy on adversarially modified AI text and 93.66% on humanized AI text that evades other detectors like GPTZero (95.60%) and Binoculars (94.40%) on the adversarial task. Their current production model (as of mid-2026) detects outputs from Fable 5 with 99.64% accuracy, though specific reasoning effort levels (e.g., High, Max) are not disclosed.

Notably, Pangram has released an open-source model (Llama-3.2-3B QLoRA) that was state-of-the-art at launch. The classifier now provides a percentage verdict instead of binary. Their Chrome extension has approximately 10,000 users, with the author estimating an upper bound of 50 million if they maintain market leadership. The company has also announced plans to open a Toronto office, which the author believes will grow faster than several established AI firms in Ontario, including Ideogram, Elevenlabs, and Cohere.

The brief also discusses strategic implications: Pangram could be an acquisition target for Thinking Machines, given the importance of text provenance for human readers. However, the author notes that detection and evasion co-evolve, potentially making humans dependent on AI detectors. False positive rates for Pangram 3.3 are reportedly below 0.5-1.5%, and sometimes below 0.05%, depending on the human-written dataset.