Pangram, known for detecting machine-written text, released two visual detection models on November 4. Pangram Image 1.0 targets still images, and Pangram Image Flash 1.0 is a smaller, faster sibling built for video.
The split is deliberate. Video detection works by analyzing frames one at a time, so speed matters more than maximum precision on any single frame. A lighter model makes scanning a full clip practical.
Pangram says Image 1.0 reaches 99.96% on an internal benchmark. Flash was measured at 98.57% on AI video against 84.2% for the next best competitor. Both figures are vendor-reported and have not been independently replicated.
The release also adds heatmap output, native video uploads and a Chrome extension. Those features push detection from a research demo into a tool that newsrooms, marketplaces and platforms can run at volume.
The underlying problem is that synthetic media no longer announces itself. Shoddy generations that are easy to spot by eye are giving way to outputs that survive casual inspection, and provenance metadata is frequently stripped during reposts.
That makes detector accuracy a competitive battleground. Pangram’s earlier image preview shipped in July to subscribers as a research preview; the move turns the company from a single-format vendor into a multi-format verification shop.
Independent testing remains the missing piece, as it does across the detection industry, where benchmarks are often internal and rivals rarely evaluate each other’s models.
