Pangram released Image 1.0 and Image Flash 1.0 on November 4, 2026, expanding its tools for identifying AI-generated visual content. The flagship image model produces a heatmap, while the smaller Flash model supports video analysis.
What the launch adds
The announcement describes image localization, native video uploads, Chrome extension integration, and API access. Pangram also lowers the minimum supported image dimensions to 256 by 256 pixels, with a reduced-confidence warning below 512 pixels.
The company reports 99.96% accuracy on its own evaluation of 1,200 synthetic images derived from 100 real photos using 12 generators. It also reports 98.7% on the DF26 video benchmark for Flash. These are vendor-reported results for particular evaluations, not independent proof of equivalent performance on every image or video a reader encounters.
What the benchmark actually covers
The original DF26 research, submitted in September, studies videos featuring one person speaking, including direct-to-camera footage, official statements, and studio interviews. It contains 271 real and 2,420 synthetic videos from seven generation systems.
The researchers report that the humans and detectors they evaluated struggled with those modern synthetic videos. That earlier study establishes the benchmark’s context, not an independent replication of Pangram’s November result. A new detector’s reported score and the original paper’s findings refer to different evaluations.
Detection is one piece of an authenticity decision
Our editorial analysis is to preserve a distinction between a detector’s classification and the broader truth of a claim. A real photograph can carry a false caption. A synthetic illustration can be openly labeled and harmless. A heatmap can help focus further inspection, but it does not establish who created the material or why.
Before adopting a detector for a consequential workflow, test a representative set of real material alongside known synthetic examples. Include the compression, cropping, screenshots, and edits that actually occur in the organization’s publishing or review process. Record false positives as carefully as missed synthetic content.
A seemingly small false-positive rate can still create a large review burden when most incoming images are authentic. The useful question is how many flagged items a team can examine and what evidence is required before rejecting a submission or accusing a person of deception.
Design a review path before automating it
Keep the original file where permitted, document the detector version, and retain the surrounding source information. Use additional evidence such as provenance records, the original publisher, and corroborating reporting. Establish an appeal or manual-review process for decisions affecting people.
Pangram’s release adds practical ways to inspect visual material at the image and frame level. Its value in a specific workflow depends on measured performance, careful interpretation, and a review process that recognizes uncertainty instead of turning a model score into a final verdict.