On November 4, 2026, Pangram Labs introduced two models on its own blog that are meant to detect AI content in images and videos. Pangram Image 1.0 checks still images, Image Flash 1.0 checks videos. Until now, the US company, which according to heise was founded in 2023, only offered detection for AI texts.
Both models run via the extended API, and Image 1.0 is also available in Pangram's browser extension. The hit rates come from the manufacturer itself and have not been independently confirmed.
What can Pangram Image 1.0 and Image Flash 1.0 do?
Image 1.0 delivers more than a verdict: a heatmap shows which areas of an image come from an AI. If an AI has inserted an object into a real photo, exactly that spot lights up, as heise describes in its example. Image Flash 1.0 is smaller and faster. The model breaks a video down into individual frames and can thus identify the sections that come from an AI.
The minimum resolution dropped from 512 by 512 to 256 by 256 pixels. If one edge is under 512 pixels, the result indicates lower reliability. The model is said to be robust against compression, scaling and image editing, but the company does not provide a figure for this.
How reliable is the detection?
Pangram states a hit rate of 99.96 percent for Image 1.0. The company measured this value on its own test set of 100 real photos and 1,200 synthetic counterparts from 12 image generators. According to Pangram, Image Flash 1.0 achieves 98.7 percent in the DF26 benchmark test and is 29 percentage points ahead of the next-best detection tool.
Pangram Image in Zahlen
Zwei Modelle für Bilder und Videos, vorgestellt am 4. November 2026
- 99,96 %Trefferquote bei Bildern im Eigentest von Pangram
- 98,7 %Trefferquote bei Videos im Vergleichstest DF26
- 256 pxMindestkante, unter 512 Pixeln sinkt die Zuverlässigkeit
Alle Werte stammen von Pangram und sind nicht unabhängig bestätigt
Quellen: Pangram, heise
The numbers are only of limited use for context. In the image test, only 100 of 1,300 images are real, and in the video test 129 of 1,529. How often a real photo is wrongly classified as AI cannot be determined from the stated values.
The DF26 study cited by Pangram shows that such tools are in demand: according to heise, people distinguish real from AI-generated videos of speaking people only slightly better than chance. In addition, heise places the launch against the background of the EU labeling requirements for AI content, which have applied since August 2026.
What does Pangram Image cost?
Pangram lists its plans only in dollars and does not state a euro price. Image checking is part of all plans: Individual includes 100 image scans per month, Professional five times that. Individual and Team start with a seven-day trial period, and permanently free access does not appear on the pricing page. Via the API, fees are charged per checked image.
What this means for you
- If you test Pangram, start with images whose origin you know: an unedited photo, the same shot after retouching, and an AI image. That way you can see whether the heatmap distinguishes retouching from AI interventions.
- Spot-check material from suppliers or clients before you pass it on. A hit is an indication, not proof, as long as independent tests are missing.
- For preview images with edges under 512 pixels, you are better off trusting the result only with caution, because Pangram itself points to lower reliability there.
Pangram introduced Image 1.0 and Image Flash 1.0 on November 4, 2026.
Sources
