On November 4, 2026, Pangram graduated its image detector from research preview to Pangram Image 1.0 and added Pangram Image Flash 1.0 for video. The launch post focuses on heatmaps that highlight likely AI regions, tougher low-resolution support, a Chrome extension, and an API for images and video.

Coffee summary

  • FACT: Image 1.0 outputs heatmaps to localize AI content inside otherwise real photos; minimum input resolution drops to 256×256 (with reduced-confidence notices below 512).
  • CLAIM (Pangram): 99.96% accuracy on a 1,200-image set of synthetic mirrors made with 12 frontier image generators from 100 real photos.
  • CLAIM (Pangram): Image Flash 1.0 scores 98.7% on DF26 — about 29 points above the next best detector in their comparison.
  • FACT: Flash analyzes uploaded video frame by frame; Image 1.0 is in Pangram’s Chrome extension; the API now covers image and video.
  • LIMIT: Pangram’s post does not publish a clear public false-positive rate on everyday real photos/videos from the open web.

What happened

Pangram’s hard test for images builds “synthetic mirrors”: take a real photo, regenerate it with many frontier generators using closely matched prompts, then ask detectors to tell them apart. For video, they started from 129 real social clips and built AI mirrors with 14 generators for 1,529 clips total. Flash is pitched especially at talking-head deepfakes, where DF26 found humans near chance and older detectors only modestly better.

They also emphasize robustness to compression, downscaling, blur, and noise — the edits social platforms and bad actors already apply.

Why it matters

Detection only helps if it fails gracefully. Heatmaps and per-frame video scores are more useful than a single yes/no, because editors can inspect the flagged region instead of trusting a black-box percentage. Still, vendor leaderboards on self-built mirror sets can look cleaner than messy real feeds.

Who should care

  • Journalists, moderators, and marketplaces screening listings, profiles, and viral clips.
  • Ordinary users who want a browser flag while scrolling shopping or social feeds.
  • Developers who need an API rather than a manual upload tool.

Limitations

The 99.96% and DF26 98.7% figures are Pangram’s. Independent audits are not included in the launch post. False-positive rates on large real-only corpora are not clearly reported there — a practical gap if you auto-block content. Generators released after the eval can degrade detectors quickly. The public launch materials highlight a Chrome extension; do not assume feature parity on every browser without checking Pangram’s current downloads.

What to do next

  1. Test Image 1.0 on a few known-real and known-AI images you already trust; note heatmap behavior, not only the score.
  2. For video, upload a short clip and review frame-level flags before acting on a single headline percentage.
  3. If you automate moderation, log false positives for two weeks before turning on hard blocks.

AIImpish Take

Pangram is shipping the right UX ideas — localization, low-res support, in-feed checks — while asking you to trust vendor benchmarks. Use it as a second opinion, publish your own error rates, and keep a human in the loop when the stakes are high.

Related on AIImpish

Sources: Pangram, Introducing Pangram Image 1.0; Pangram product site.