Pangram announced Pangram Image 1.0 on November 4, 2026. The company calls it its flagship model for detecting images made by frontier AI image generators. It released a smaller, faster companion, Pangram Image Flash 1.0, focused on AI-generated video.
Pangram reports that Image 1.0 reached 99.96% accuracy on an internal set of 1,200 synthetic images. The set was built by mirroring 100 real photos across 12 frontier image generators, which the post does not name. Pangram says Image Flash 1.0 scores 98.7% on the DF26 benchmark, 29 points above the next-best detector. It also compared detectors on 1,529 videos: 129 real social-media clips plus synthetic copies from 14 video generators. Those results, like the academic benchmark results, appear only in charts.
Image 1.0 outputs a heatmap that shows where AI-generated content appears within an image, not just a single verdict. Pangram says the model is trained to stay robust to compression, downscaling, blur, noise and other edits. The minimum input size has dropped from 512x512 to 256x256 pixels, and results below 512 pixels come with a reduced-confidence disclaimer. Uploaded videos are analyzed frame by frame, with flagged segments marked.
The detector is built into Pangram's Chrome extension, which flags AI content while users browse social media, shopping sites and rental listings. A Firefox version is also linked. Pangram's API now supports both image and video detection. The announcement offers a free trial but does not list prices.
For trust-and-safety and moderation teams, the main changes are region-level output and video support through the same API. The accuracy figures come from Pangram's own evaluations, so teams should test the detector on their own data before relying on it.