Pangram Image 1.0 Achieves 99.96% AI Detection Accuracy

Pangram officially launched Image 1.0 and Image Flash 1.0 on November 4, 2026, claiming state-of-the-art performance in detecting synthetic images and deepfake videos. The company reports 99.96% accuracy on a test set of 1,200 AI-generated images from 12 frontier models, and 98.7% accuracy on the DF26 video benchmark, a 29-point improvement over the next best commercial detector. The release follows months of community testing of Pangram's research preview and comes one week after an NPR peer-reviewed study validated the company's text detection capabilities.

Record-Breaking Detection Performance

Pangram evaluated Image 1.0 against a purpose-built benchmark of 1,200 synthetic images created by mirroring 100 real photographs across 12 state-of-the-art image generators. Each synthetic version used prompts closely matching the original photo to create difficult detection scenarios. The 99.96% accuracy represents near-perfect classification, with the company comparing its model against several commercial and academic AI image detectors on the same evaluation set.

For video detection, Image Flash 1.0 achieved 98.7% accuracy on the DF26 academic benchmark, which previous research showed stumps human reviewers who perform barely better than random guessing. The company created an additional evaluation set of 1,529 videos by generating synthetic mirrors of 129 real social media videos using 14 different frontier video generators, seeding each with the first frame of the real video and a vision-language model description of the action.

Expanded Resolution Support and Visualizations

The production release lowers the minimum supported image size from 512x512 pixels to 256x256 pixels, addressing a key limitation of the research preview. Pangram reports improved accuracy at lower resolutions, though results below 512 pixels include a reduced confidence disclaimer. This expansion matters for real-world deployments where social media platforms compress and downscale user uploads.

Image 1.0 now includes heatmap visualizations that highlight specific regions of an image flagged as AI-generated. The company demonstrated this capability with a photograph containing embedded AI content, showing the model's ability to localize synthetic elements within otherwise authentic images. The system is trained to detect common evasion techniques including blur, noise, and compression artifacts that malicious actors use to fool detection systems.

Browser Extensions and API Access

Pangram integrated Image 1.0 into Chrome and Firefox extensions that flag AI-generated content as users browse social media, online shopping sites, rental listings, and other platforms. The extensions operate in real time, analyzing images without requiring manual uploads. For video analysis, the platform now supports native video upload with frame-by-frame analysis that identifies which segments contain AI-generated content.

The company updated its API to support both image and video detection, enabling third-party developers to integrate the models into their own applications. Pangram previously released detection capabilities as a research preview to gather community feedback on edge cases and failure modes, a testing approach the company credits with improving the production model's robustness.

Release Date and Availability

Pangram Image 1.0 and Image Flash 1.0 officially launched on November 4, 2026. The models are available through the company's web platform, browser extensions for Chrome and Firefox, and a newly updated API. Pricing for API access was not disclosed in the announcement, though the company emphasized making detection capabilities accessible to the broader community.

What This Means

Pangram's 99.96% accuracy claim, if validated by independent testing, represents a significant improvement in AI content detection at a time when synthetic media proliferates across social platforms. The 29-point gap over competitors on video detection addresses a critical security concern as deepfake videos become increasingly difficult for humans to identify. Browser extension deployment lowers the barrier for everyday users to verify image authenticity, though the effectiveness of any detection system ultimately depends on ongoing adaptation as generative models evolve. The company's decision to train against common evasion techniques suggests awareness that detection is an arms race rather than a solved problem.