Pangram launches Image 1.0 and Flash 1.0 to detect AI in images and video
Pangram’s new detectors map AI content inside images and check video frame by frame; all accuracy figures are the company’s own.
Published
Last updated / Fact-checked on
AI-detection company Pangram released two models on Nov. 4, 2026: Pangram Image 1.0, which outputs a heatmap showing where AI content sits inside an image, and Pangram Image Flash 1.0, a smaller, faster model aimed at video. Both are now available through Pangram’s API, and Image 1.0 powers the company’s Chrome extension, according to a Pangram blog post.
Photo: PixxlTeufel / Pixabay
At a glance
| Pangram Image 1.0 | Pangram Image Flash 1.0 | |
|---|---|---|
| Role | Flagship image detector | Smaller, faster model aimed at video |
| Output | Heatmap localizing AI content in an image | Frame-by-frame analysis of uploaded video |
| Headline result (self-reported) | 99.96% accuracy on 1,200 synthetic “mirrors” of 100 real photos (12 generators) | 98.7% on the DF26 deepfake benchmark |
| Where to use it | Chrome extension; API | Native video uploads; API |
All figures are Pangram’s own and have not been independently verified.
Pangram announced the models in a blog post titled “Introducing Pangram Image 1.0,” written by founding computer vision engineer Isaac Kasahara. The release follows a research preview of Pangram Image that, according to the post, the company put out “a few months ago.”
The launch matters because synthetic images and videos now turn up on social feeds, shopping pages and rental listings. Pangram’s performance claims, however, come from its own testing and have not been independently checked.
What Image 1.0 does
According to Pangram’s post, Image 1.0 produces a heatmap that localizes AI content within an image. The company’s example is an AI-generated element placed inside an otherwise real photograph. Pangram says it trained the model to hold up against edits, compression, downscaling, blur and noise.
The minimum input size drops from 512x512 to 256x256 pixels, the post states. Results for images below 512 pixels come with a reduced-confidence disclaimer.
To test the model, Pangram says it built 1,200 synthetic “mirrors” of 100 real photos using 12 frontier image generators. It reports 99.96% accuracy on that set. Pangram also says it compared the model on five academic benchmarks: X-AIGD, Mirage-Test, AIGenImages2026, Synthbuster + RAISE, and Community Forensics. These are Pangram’s figures and have not been independently verified.
What Flash 1.0 does for video
Pangram says Flash 1.0 scores 98.7% on DF26, an academic deepfake benchmark of talking-person videos, and calls that “a 29-point leap over the next best detector.” Pangram cites DF26 as showing that humans are barely better than chance at telling real videos of people talking from AI-generated ones.
Pangram also reports state-of-the-art results for Flash 1.0 on two other benchmarks, AIGVDBench and Ivy-Fake. It describes an industry comparison of 1,529 videos: 129 real social media videos plus “mirrors” made with 14 video generators, each seeded with the real video’s first frame. Pangram says it tested against “multiple major AI video detection services” but does not name them in the post. These results are also self-reported.
Video uploads are now supported natively, according to the announcement. The analysis runs frame by frame and shows which parts of a clip are AI-generated.
Where users will see it
Pangram says Image 1.0 is built into its Chrome extension, which flags AI-generated content while users browse social media, shopping sites, rental listings and other sites. Pangram’s API now handles both image and video detection, the post states.
What to watch
- Independent testing. Every accuracy figure here comes from Pangram. Third-party evaluations on everyday content would show whether the 99.96% and 98.7% numbers hold up.
- Small and heavily edited images. Pangram warns that results under 512 pixels carry reduced confidence. Watch how the heatmap handles reposted, recompressed images.
- Named comparisons. Pangram did not name the video detection services it beat.
- Practical use. Treat any detector’s verdict as one signal, not proof.
More coverage is on our Artificial Intelligence hub.
How we reported this
We based this article on Pangram’s Nov. 4, 2026 blog post, which we read directly. All performance numbers are Pangram’s own, self-reported results and have not been independently verified. We kept Image 1.0 and Flash 1.0 results separate because they cover different tasks. We left out competitor scores, pricing, model architecture, DF26 authorship details and extension install numbers because the post does not give them or we could not verify them.
Published and fact-checked November 4, 2026.
Sources
- Isaac Kasahara, “Introducing Pangram Image 1.0,” Pangram blog, Nov. 4, 2026: https://www.pangram.com/blog/introducing-pangram-image-1-0
DailyMech used AI tools to help research and draft this article; it was fact-checked by our editors.