Pangram 4 jumps ahead of its predecessor in size and accuracy. At over six times larger than Pangram 3, it drops false positives to 0.0041%—that’s one in 24,000 documents—and cuts false negatives from 1.99% down to 0.3396% on new challenge sets. The model flags AI-edited text and mixed human–AI passages in a single pass and spots AI involvement in humanized output with 98.83% success across 13 commercial tools. It holds detection rates under 0.7% false negatives for every major model family, regardless of release date, proving its strength against frontier systems and adversarial prompts. Pangram 3 remains available until September 30, 2026, but defaults will route to the new model only if you specify Pangram 4. Credit pricing shifts from rounding scans up to 1,000 words to charging per 100 words. Individual plans now cover 300,000 words and 100 image scans per month; Pro plans rise to 1.5 million words and 500 image scans—all without raising prices. API calls go for $0.05 per 100 words, a boost for short docs and up to ten times costlier for long-form text. For full test results, training info, and limitations, you can read the Pangram 4 technical blog or model card. Max Spero, the co-founder and CEO with a background at Google, Two Sigma, Yelp and Nuro, signed off on this release.