Pangram Image 1.0: the new detector for AI images and video
Pangram launched two models on November 4, 2026 that detect AI-generated images and video. What its numbers mean, what they leave out, what it costs, and what changes if you publish AI visuals.
Frank HoubreFounder of Imaginode
6 min read
What Pangram announced on November 4
New York startup Pangram launched two detection models on November 4, 2026: Pangram Image 1.0 for still images, and the lighter Pangram Image Flash 1.0 for video. If you publish AI-generated visuals, your clients and readers now have one more way to spot them.
Pangram's post follows a research preview that The Next Web dates to July, when the company raised $9 million. Until now Pangram was mostly known for catching AI-written text, and its tool runs on Substack.
You'll notice the heatmap first. On top of the score, Image 1.0 shows where the AI sits in the picture. In the launch example, a real photo of a restaurant, only the wall mural lights up. For video, you upload a file and it gets checked frame by frame. The model is also in the Chrome extension, which flags AI content as you scroll social media, online stores or rental listings.
The minimum image size drops from 512 x 512 to 256 x 256 pixels, with a reduced-confidence warning below 512. So even a thumbnail can be checked.
The claimed numbers, and how far to trust them
Pangram claims it catches 99.2% of AI images and wrongly flags 0.27% of real ones. Those scores come from the company, and I found no independent check as of November 5.
Run those rates and you get about 8 AI images missed per 1,000, and about 27 real photos wrongly flagged per 10,000. On a site checking millions of images, that's a lot of real photos under suspicion.
The main test in the post was built by Pangram. It took 100 real photos and had 12 models generate close copies, for 1,200 AI images. Claimed result: 99.96% balanced accuracy, ahead of TruthScan (96.33%), AI or Not (91.38%) and Sightengine (88.33%). The chart notes that those rivals' scores date from October. And 100 real photos isn't many: a single mistake is already 1% of the set.
The list of models tested is the part to read: FLUX.3, Seedream 5.0 Pro, Nano Banana 2.1, Ideogram 4.5, MAI-Image 2.6, Midjourney 8.2, GPT Image 2.5 Flare and Sunburst, Recraft V4.1 Flash, Grok Imagine 2.0 and others. Several are in Imaginode's catalog. GPT Image 2.6 came out the same day and isn't on the list, so there's no number for it.
Video: 98.7% or 98.57%? Two different tests
Pangram Image Flash 1.0 scores 98.7% on DF26, an academic benchmark published in September, and 98.57% on an in-house test of 1,529 videos. Both numbers are going around because they measure different things.
DF26 has 271 real videos and 2,420 synthetic ones from seven recent models. Its authors found that people and existing detectors alike do little better than chance. Pangram claims a 29-point lead over the next detector. The dataset is external, but the score is the one Pangram reports for itself. DF26 also only contains people talking: statements, interviews, direct to camera. It tells you nothing about a film scene with a dragon.
The in-house test covers 14 generators, including Seedance 2.5, Veo 3.1, Wan 3.0, FLUX.3 Video and Grok Imagine Video 1.5. Each fake starts from the first frame of a real social media video. Pangram gives itself 98.57%, against 84.2% for TruthScan, 81.8% for Sightengine and 65.2% for AI or Not.
Compression, blur, screenshots: what the example shows
Pangram says it trains on compression, blur and noise, and it publishes one AI image run through ten transformations: the score stays between 92.9% and 98.5%.
In that example the original scores 98.5%, a heavily compressed JPEG 96.8%, a screenshot 96.1%, and a double re-encode 92.9%. Keep in mind that one image is an illustration and nothing more. The image comparison was also run with metadata stripped, so the model isn't just reading the file's labels.
Watermarks showed their limits this summer. In July, Reuters tested Meta's detector, which relies on an invisible watermark: of 40 cropped images, it missed 55%, The Next Web reports. Pangram analyzes the image itself. Whether that holds up better over time, model after model, will take outside testing to find out.
If your plan is to hide that an image was generated, that plan ages badly. On the screenshot in that example, the July preview gave 67.5%. Version 1.0 gives 96.1%, four months later.
What it costs
The Individual plan is $20 a month after a 7-day trial, with up to 300,000 words or 600 images a month. On the API, Image 1.0 costs $0.25 per image.
On the API, a one-minute video costs $3 and a five-minute short film costs $15. That pricing is aimed at platforms and teams more than solo creators.
Pangram plans and API rates checked on November 5, 2026
| Plan or use | Price | What you get |
|---|---|---|
| Individual | $20 a month | 300,000 words or 600 images a month, $10 in API credits, 7-day trial |
| Professional | $65 a month | 5x more scans, $200 in API credits |
| API Image 1.0 | $0.25 per image | The main model for still images |
| API Image Flash 1.0 | $0.05 per image | The lighter model |
| API video | $0.05 per second | Frame-by-frame analysis |
The case that can catch you out: photos edited in a chatbot
A real photo edited in ChatGPT or Gemini can get flagged as AI in its entirety, because those tools often redraw far more than the area you asked about.
Pangram showed this in a September 28 post. The author asks for ducks in the pond of a garden photo. The ducks show up, and so do changes to passers-by's faces, some branches and other details well outside the pond. The post warns that detectors may then mark the whole image.
Pangram's advice is to keep the original and paste back only the generated area. On Imaginode, the Retouch node does that: one layer with your original photo, one with the edited version, and you erase just the area you want so the new part shows through. Everything else stays your own file, pixel for pixel. I haven't run that composite through Pangram's detector: the goal here is to preserve your photo, and I won't promise a score.
What I'd do if you publish AI visuals
Say you use AI, keep your original files, and never treat a detector score as proof.
For client work, put the method in the quote. I explain why in the guide on pricing AI creative work, and the guide on rights to AI images covers what you can actually hand over. A client with the extension may run into the flag on their own, so it's better they read it in your quote first.
For a store or a listing, tell buyers the visual is an illustration. Those are exactly the pages the extension targets. For video, check each platform's rules too: Snap announced on July 31 that wholly AI-generated videos are no longer recommended on Spotlight.
And if you're the one wrongly accused? Keep your original photos, your prompts and your working steps. In August, Pangram's text tool rated a Berkeley professor's op-ed as partly AI, and she disputed the result. A 0.27% rate says nothing about your specific image: if someone quotes a score at you, ask for the heatmap and show your files.
How we analysed this
Written on November 5, 2026 from Pangram's November 4 post (by Isaac Kasahara), its September 28 post on AI editing, its pricing page, The Next Web's November 4 article, the DF26 abstract on arXiv and Snap's July 31 announcement. Screenshots come from Pangram's pages and were taken on November 5, 2026. The 99.2% and 0.27% figures come from Pangram's X account and The Next Web: the official post doesn't break them down.
We did not test Pangram, and no account was created for this article. All scores are the ones Pangram publishes. Disclosure: Imaginode, Frank's product, generates images and video with AI, so its users' work is directly affected by these detectors.
Sources