Pangram releases Image 1.0, claiming 99.96% accuracy on AI image detection, plus Image Flash 1.0 for video

Pangram has released Pangram Image 1.0, a flagship AI image detection model that the company claims achieves 99.96% accuracy on its internal evaluation set. It also released Pangram Image Flash 1.0, a smaller, faster model that powers its AI video detection and, according to Pangram, scores 98.7% on the DF26 benchmark.

Pangram Image 1.0 follows a research preview of Pangram Image released a few months ago. Pricing, parameter counts and context window are not yet disclosed.

Pangram Image 1.0

Pangram says the model delivers state-of-the-art detection of frontier AI image generators. New features include:

• Heatmap output that localizes where AI content appears in an image, including a real photograph that contains an AI-generated element.
• Robustness to edits, compression and low resolutions. Pangram says it trains specifically against blur, noise, downscaling and other artifacts that social platforms or bad actors introduce.
• Lower minimum input size. The research preview required at least 512x512. The new limit is 256x256. Results below 512 carry a reduced-confidence disclaimer.

Evaluation

Pangram tested the model against several commercial and academic detectors. For its internal set, it created synthetic mirrors of 100 real photos using 12 frontier image generators, with prompts closely matched to each original, for a total of 1,200 synthetic images. Pangram says Image 1.0 reaches 99.96% accuracy on this set.

Pangram also compared the model on academic benchmarks: X-AIGD, Mirage-Test, AIGenImages2026, Synthbuster + RAISE, and Community Forensics.

Pangram Image Flash 1.0

Image Flash 1.0 is the model behind Pangram's video detection. Pangram cites DF26, an academic study that it says found humans barely better than random at telling real videos of people talking from AI-generated ones, with modern detectors not much better. On DF26, Pangram says Flash scores 98.7%, a 29-point leap over the next best detector.

Industry comparison

Pangram built a test set from 129 real social media videos. It generated AI mirrors using 14 frontier video generators, 100 videos each, sampled from the 129 real videos. Each generator was seeded with the first real frame of the source video and prompted with a VLM-generated description of the video's content. The total evaluation set was 1,529 videos. Pangram compared its model against multiple major AI video detection services on this set.

Because each real video and its mirror share the same scene and subject, Pangram says the design prevents the model from developing bias toward labeling certain content types as AI.

Pangram also reports state-of-the-art results on academic video benchmarks AIGVDBench, DF26 and Ivy-Fake, which it says together cover almost every frontier video generation model.

Product integration

• Native video upload: Image Flash 1.0 analyzes uploaded video frame by frame and reports which parts are AI-detected.
• Chrome extension: Pangram Image 1.0 now flags AI-generated content while users browse social media, shopping and rental listings, and other sites.
• API: Pangram updated its API to support both image and video detection.

What this means

The accuracy figures come from Pangram's own test sets and have not been independently verified. The 99.96% result applies to a 1,200-image set built from 100 real photos, and the video comparison uses 1,529 videos built from 129 real clips. Both are paired real-versus-synthetic designs, which test discrimination between matched content rather than detection across arbitrary real-world material.

The practical changes are the heatmap output, which shows where in an image AI content sits, the 256x256 minimum input size, and frame-by-frame video analysis. Pricing for the new models has not been disclosed.