Pangram Image 1.0 is Pangram Labs' November 4, 2026 release for spotting AI made or AI edited images, with a faster sibling model, Pangram Image Flash 1.0, aimed at video frame by frame. The flagship model can draw a heatmap that marks where AI content sits inside a photo. Flash is what runs when you upload a video in the web app. Both ship in the product, the API and (for Image 1.0) the Chrome extension. Accuracy numbers on the launch post are mostly Pangram's own tests, so treat them as vendor claims until someone else reproduces them on public sets.
This guide covers what the two models do, how the DF26 video benchmark frames the problem, what free and paid access look like on November 6, 2026, and how Pangram Image fits next to watermark and provenance checks. We read Pangram's launch post, pricing page and API page, plus the DF26 paper abstract. We did not create a Pangram account or run a first hand image or video scan. For text detection and plan math, see is Pangram Labs free. For the wider media stack, see deepfake detection and how to check if a video is AI generated.
What Pangram Image 1.0 is
Pangram started as an AI text detector. Image support arrived as a research preview months earlier. On November 4, 2026 the company promoted that preview to Pangram Image 1.0 and added Pangram Image Flash 1.0 for video (launch post). The same day, independent coverage on The Next Web and other outlets repeated the claims and noted that the headline accuracy figures are self reported.
In product terms the release adds four concrete surfaces:
- Still image scans with Image 1.0, including a localization heatmap for mixed photos (for example a real restaurant wall with an AI mural pasted in).
- Native video upload that runs Flash frame by frame and shows which stretches look synthetic.
- Chrome extension support for Image 1.0 while you browse social feeds, shopping pages and rental listings.
- API endpoints for both image models and for video billed by the second (API page).
Image 1.0 is the careful model. Flash is the smaller, faster one Pangram says is tuned for frontier video generators. You do not pick Flash for a still photo in the marketing copy; you pick Image 1.0 for photos and Flash for video.
How it differs from watermarks and credentials
Pangram Image is a classifier. It guesses from visual patterns whether pixels look AI generated or AI edited. That is a different job from:
- SynthID Detector, which only finds Google DeepMind's invisible watermark when a participating tool (Google, OpenAI, NVIDIA, Kakao, Apple for images) embedded it.
- Content Credentials Verify, which reads a signed C2PA history when the metadata survived upload and reencoding.
- OpenAI's verification tool, which checks OpenAI image and audio signals, not arbitrary video files.
A positive watermark or credential is strong evidence that a participating tool touched the file. A negative result proves almost nothing. A Pangram Image score is a lead for unlabeled media, with the usual classifier limits: new generators, heavy compression and partial edits can move the score the wrong way. Pair classifier output with source checks. Our deepfake detection guide walks through that stack for images; the video checklist does the same for clips.
Vendor scores, and why they need context
Pangram's launch post lists several figures. Quote them as vendor results, not as independent audits.
Image 1.0 (vendor tests). Pangram built "synthetic mirrors" of 100 real photos with 12 frontier image generators (1,200 fakes total) and reports 99.96% accuracy for Image 1.0 on that set. Separately, coverage of Pangram's launch materials cites a 99.2% catch rate on AI images with a 0.27% false positive rate on real photos (vendor figures, not in the main blog table). It also claims better resistance to compression, blur, noise and lower resolutions, and it drops the minimum input size from 512 by 512 to 256 by 256, with a reduced confidence warning below 512.
Image Flash 1.0 on DF26 (vendor claim on a public benchmark). DF26 is an academic video set of 2,691 clips: 271 real and 2,420 synthetic public speaking videos from seven modern generators (arXiv:2609.07369, September 7, 2026). The authors report that both people and prior detectors were close to chance on that set. Pangram says Flash scores 98.7% on DF26 and leads the next detector by about 29 points. That is Pangram's figure on DF26, not a third party table we rechecked.
Industry style video mirrors. Pangram also describes 129 real social videos mirrored with 14 generators (1,529 clips) and says Flash beat "multiple major AI video detection services" on that set. The public post does not name those services or publish a full score table we could copy, so we do not invent ranks.
Press coverage on November 4 and 5 (The Next Web, Unite.AI, Heise) repeated these numbers and flagged the self reported nature of the image mirror test. Until independent labs publish matched runs, treat the SOTA claim as marketing backed by a vendor evaluation, not as a settled leaderboard.
DF26 in one paragraph
DF26 matters even if you never trust a vendor score. The paper builds matched real and fake public speaking clips (direct to camera, official statements, studio interviews) and finds that humans and older detectors barely beat chance. That is why a product that claims a large jump on DF26 draws attention, and why you still need source checks for high stakes clips. Access to the dataset is controlled through a request form on Hugging Face; training on the closed source subset is restricted. Pangram's own score is the main public Flash claim we have seen as of November 6, 2026.
Free limits, seat plans and API prices
Access sits on the same Pangram account ladder as text detection. Details below were rechecked on Pangram pages on November 6, 2026.
| Path | What you get for images and video | Cost signal (Nov 6, 2026) |
|---|---|---|
| Free web plan | 3 image detection scans per day, separate from the daily text credit bucket | 0 dollars, no card (free account help; is Pangram Labs free) |
| Individual | 100 image scans per month, plus higher text caps | 20 dollars a month, or 15 dollars a month billed annually (pricing) |
| Professional | 500 image scans per month | 65 dollars a month, or 45 dollars a month billed annually |
| Developer API | Prepaid credits; Image 1.0, Flash and video billed separately | Image 1.0: 0.25 dollars per image; Image Flash 1.0: 0.05 dollars per image; video: 0.05 dollars per second (API) |
The free help article focuses on the daily image scan count; it does not spell out a separate free video upload quota in the same clear way. If you need video on Free, confirm in the live app before you rely on it. Paid seat plans advertise monthly image allowances. API buyers pay per image or per second of video from a credit balance (purchase cards on the API page start around 25 dollars, with smaller research paths described elsewhere). Individual includes about 10 dollars of monthly API credits and Professional about 200 dollars, per the live pricing page.
Chrome extension users still need a Pangram login for product features; the extension now runs Image 1.0 on feed images according to the launch post.
How to use it without overtrusting a score
- Start with provenance and watermarks when the file might come from Google, OpenAI, Apple or another SynthID partner. Use SynthID Detector and Content Credentials Verify first.
- Upload the highest quality original you can find. Screenshots, social recompression and heavy crops weaken both watermarks and classifiers. Pangram says Image 1.0 is trained for compression and low resolution, but a reduced confidence warning still appears below 512 pixels.
- Read the heatmap on mixed images. A mural, product insert or face swap inside an otherwise real photo is exactly the case the launch demo highlights.
- For video, use Flash results as a timeline, not a verdict. Frame level flags help you zoom in; they do not replace calling a person back on a known number when the clip asks for money or access.
- Keep a second classifier opinion for high stakes work. Hive offers a free Chrome extension with no Hive login. Reality Defender has a free developer API for images and audio (video is paid); see is Reality Defender free.
Pangram Image vs Hive and Reality Defender
| Need | Prefer |
|---|---|
| Quick browse time classifier with no account | Hive's free Chrome extension |
| Heatmap localization on a still photo inside Pangram's product | Pangram Image 1.0 |
| Consumer web upload for video with Pangram's Flash model | Pangram Image Flash 1.0 (confirm plan limits) |
| Enterprise live call and meeting detection | Reality Defender (sales quote; free API excludes video) |
| Confirm Google or partner watermark | SynthID Detector |
| Confirm signed edit history | Content Credentials Verify |
Pangram's edge in this launch is the combination of a consumer familiar brand, heatmap localization, native video upload and public API rates for image and video. Hive remains the easiest no login browser check. Reality Defender remains the enterprise stack for live media and security workflows. None of them replace provenance when a credential or watermark is present.
Who should care about this release
Journalists, moderators, landlords screening listing photos, and anyone who already uses Pangram for text now have an in product path for images and (on supported plans) video. API teams can price Image 1.0 at 0.25 dollars per image and Flash or video at 0.05 dollars per image or per second without a sales call for the prepaid tier. Researchers watching DF26 finally have a commercial model that claims a large gap on that hard public speaking set, which is newsworthy even before independent replication.
Skip Pangram Image as your only check when the stakes are legal, medical or financial. Combine it with source investigation, and prefer watermark or credential evidence when it exists.
FAQ
What is Pangram Image 1.0?
It is Pangram Labs' November 4, 2026 AI image detection model. It scores whether a still image looks AI generated or AI edited and can output a heatmap that marks the AI regions. A smaller model, Pangram Image Flash 1.0, handles video frame by frame.
Is Pangram Image free?
Partly. The Free Pangram plan includes 3 image detection scans per day with no card required. Heavier image volume needs Individual (100 scans a month) or Professional (500). API image and video calls use prepaid credits. Full plan math is in is Pangram Labs free.
How is Pangram Image Flash different from Image 1.0?
Image 1.0 is the flagship still image model with heatmap localization. Flash is smaller and faster and is the model Pangram highlights for frontier AI video, including native video upload analysis.
Did Pangram Image beat every other detector on DF26?
Pangram reports 98.7% for Flash on DF26 and a large lead over the next detector it measured. That figure comes from Pangram's launch materials. We have not seen a full independent leaderboard republishing the same comparison as of November 6, 2026.
Can Pangram Image replace SynthID Detector?
No. SynthID finds a watermark only when a partner tool embedded it. Pangram Image guesses from pixels when no watermark exists. Use both: watermarks first when available, classifiers for unlabeled files.
Does the Chrome extension detect AI video?
The launch post says Image 1.0 is integrated into the Chrome extension for feed images. Native video upload in the web app is the surface described for Flash. Confirm live extension behavior for video on the page you care about; do not assume every feed video is scanned the same way as a still.

