Pangram's new detector claims 98.7% on AI video
Pangram also launched Image 1.0 for stills, which the company says hits 99.96% accuracy on a 1,200-image set it built itself.
Key takeaways
- Pangram Image 1.0 and Pangram Image Flash 1.0 launched on November 4, 2026; Image 1.0 reads still images and Flash 1.0 checks video frame by frame.
- Pangram says Flash 1.0 scores 98.7% on the DF26 academic benchmark, a 29-point lead over the next-best detector.
- Pangram says Image 1.0 reaches 99.96% accuracy on a set of 1,200 synthetic images the company built from 100 real photos with 12 generators.
- Pangram's launch post on X, as reported by other outlets, gives different image figures: 99.2% of AI images caught with a 0.27% false-positive rate.
Pangram launched two AI detection models on November 4, 2026, one for still images and one for video. The company announced Pangram Image 1.0 and the smaller Pangram Image Flash 1.0 in a blog post written by founding computer vision engineer Isaac Kasahara.
Image 1.0 reads photos and returns a heatmap marking which parts of a picture are synthetic. Flash 1.0 checks video frame by frame. Both now run through Pangram's API for images and video, and Image 1.0 is built into the company's Chrome extension, which flags AI content on social feeds, shopping sites and rental listings. Video upload is native in the dashboard now. The minimum image input drops from 512x512 pixels to 256x256, and anything under 512 comes back with a reduced-confidence disclaimer.
Pangram is a New York startup best known for AI text detection, which already runs on Substack. It raised $9 million in a round led by Menlo Ventures in July 2026, when it first showed image detection as a research preview.
The accuracy figures are Pangram's own. The company says Image 1.0 hits 99.96% on a set of 1,200 synthetic images it built by making AI mirrors of 100 real photos with 12 frontier generators. For video, Pangram says Flash 1.0 scores 98.7% on DF26, an academic benchmark of 2,691 talking-head clips published in September, a 29-point lead over the next-best detector.
Pangram also ran a separate comparison over 1,529 videos, made by mirroring 129 real social media clips with 14 video generators. Digital Trends, quoting Pangram's launch post on X, reports that post put Flash at 98.57% against TruthScan at 84.2%, the next-best competitor, with Sightengine at 81.8%.
The X post and the blog do not agree on the image numbers. Unite.AI and The Next Web report that the X post claims 99.2% of AI images caught with a 0.27% false-positive rate, while the blog reports 99.96% on its own evaluation set. Pangram has not explained the difference, and the X post itself could not be opened for this story because X refused the request.
No outside group has tested either model, so every figure here is a company claim. For creators shipping AI video, a detector that reliably flags synthetic frames could matter as platforms tighten their rules; The Next Web notes Snapchat has stopped recommending AI videos on Spotlight.
Pangram Image 1.0 and Image Flash 1.0 are live now in the web app and API, with Image 1.0 in the Chrome extension; the launch post lists no prices.
Sources
- pangram.com - launch post with model names, the 99.96% image set, the 98.7% DF26 claim and the new features
- digitaltrends.com - quotes the company's X post and names the TruthScan and Sightengine figures
- thenextweb.com - DF26 benchmark details, funding and the 99.2% / 0.27% X-post figures
- unite.ai - the 1,529-video industry comparison and the X-post figures