Pangram, the company known for detecting AI-written text, released two models for images and video on 4 November. It says Pangram Image 1.0 is “The state-of-the-art in both AI image and AI video detection”, and that a smaller model, Pangram Image Flash 1.0, scores 98.7% on the DF26 video benchmark, “a 29-point leap over the next best detector.”
On Pangram’s own figures, that puts the next best detector at about 69.7% on DF26, by our arithmetic. The benchmark results are the company’s claims; we have not seen independent tests.
Business Pill · Verification Cost
An evergreen Business Engineer explainer on verification cost, what it takes to check that something is what it claims to be. It teaches the general idea only and does not discuss Pangram.
The key insight: Pangram says its new Image Flash 1.0 model scores 98.7% on the DF26 video benchmark, 29 points ahead of the next best detector, which on its figures scored about 69.7%, by our arithmetic; every accuracy figure is the company’s own.
In a blog post by Isaac Kasahara, Pangram said it released Pangram Image as a research preview a few months ago and is now launching version 1.0. Its list of what is new includes heatmap output “for localizing exactly where AI appears in images”, robustness to edits, compression and low resolutions, frame-by-frame analysis of AI-edited videos, Chrome extension support and API support for image and video.
Pangram dated the research preview 29 July 2026, 98 days before the 1.0 release, by our count. It also lowered the minimum image size from 512×512 to 256×256, saying the new model has improved accuracy at lower resolutions; results below 512 come with a reduced-confidence disclaimer.
DF26 video benchmark scores as Pangram states them: 98.7% for Pangram Image Flash 1.0, which Pangram calls a 29-point lead over the next best detector; the ~69.7% for that detector is our arithmetic from Pangram’s figures. Company claims, not independently verified.
What Was Posted on X
Pangram is expanding into image and video detection.
Andrew Curran’s post on Pangram’s expansion into image and video detection, with a screenshot of a post by Max Spero on building a model robust to phone-camera AI postprocessing.
How Pangram Says It Tested
For images, Pangram says it created synthetic mirrors of 100 real photos with 12 frontier AI image generators, 1,200 synthetic images in all, and that “Our model achieves 99.96% accuracy on this evaluation set.” It also compares the model on five academic benchmarks, whose results it shows as charts.
For video, Pangram says it took 129 real social media videos and generated AI mirrors with 14 frontier video generators, 100 videos each, and compared its model against multiple major AI video detection services across 1,529 videos. It does not name the services in the text we read.
On DF26, an academic study Pangram cites, the company writes that humans were barely better than random at telling real videos of people talking from AI ones, and that “Modern detectors were not much better.”
What Was Posted on X
Andrew Curran posted on X at 16:39 UTC that “Pangram is expanding into image and video detection.” His post shows a screenshot of a post by Max Spero, whom Pangram lists as the author of its Pangram 4 announcement, saying: “Most phone cameras, including iPhones, do some level of AI postprocessing to add detail to real photos.”
Spero’s post continues: “We had to build a model that’s robust to this postprocessing while understanding the difference between AI edits and fully AI-generated visuals.”
Connecting the Coverage
Pangram’s text detector appeared in our coverage of a University of Washington study: AI-Written Text Reaches 24% of Web Pages, Up From Near Zero. Its authors found the open Pangram model had fewer false positives than Binoculars-Falcon on .gov and .edu domains but more on .com domains. On text, Pangram’s case studies dated 2 November say it detects GPT-6.1 Sol in 98.78% of its tested examples and Gemini 4 Argon in 99.02%.
The Structural Read
As we read the announcement, Pangram is selling localisation as much as a verdict: the heatmap marks where AI appears in an image, and the video model analyses footage frame by frame.
As we read Pangram’s method, its tests pair each real photo or video with AI mirrors of the same scene, so a detector cannot score well by flagging a type of content.
Pangram, 4 Nov 2026
“We specifically train to detect these kinds of transformations, and we have found our model to be robust to these edits.”
Three Implications
FOR PLATFORMS Pangram says it built Image 1.0 into its Chrome extension to flag AI-generated content while users scroll social media, shop online or browse rental listings.
FOR DEVELOPERS Pangram says it updated its API to support both image and video detection; the post gives no prices.
FOR RIVAL DETECTORS Pangram says it compared its video model against multiple major AI video detection services across 1,529 videos, without naming them in the text we read.
The framework says: “Each moat is built by a different data pipeline.”
As we read it, Pangram describes its own pipeline: synthetic mirrors of real photos and videos made with 12 image and 14 video generators. The framework predates this release and does not discuss Pangram.
Every accuracy figure here is Pangram’s own, from tests it designed; we found no independent evaluation of Pangram Image 1.0 or Image Flash 1.0. The text we read does not name the detectors or services it compared against, and the benchmark scores shown as charts were not read. Pangram did not give prices for the image and video API in the post.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Pangram released Pangram Image 1.0 and Image Flash 1.0 on 4 November, claiming 99.96% accuracy on its own 1,200-image test and 98.7% on the DF26 video benchmark, 29 points ahead of the next best detector; on its figures that detector scored about 69.7%, by our arithmetic.
A note on sourcing. All accuracy figures are Pangram’s own claims from its 4 November 2026 blog post and its own tests; we have not seen independent evaluations, and the benchmark charts on the page were not read. The ~69.7% and the 98-day count are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
