99.2% Detection and a 0.27% Error Rate. Read What They Tested On.

Pangram shipped two computer vision models on November 4. Pangram Image 1.0 detects AI-generated images. Pangram Image Flash 1.0 is the smaller, faster sibling, and its job is to walk through an AI-generated video frame by frame and say which parts came from a generator.
The launch adds heatmap output, native video upload, a Chrome extension, and API access for both images and video. On X, the company said Image 1.0 identifies 99.2 percent of AI-generated images while holding a false-positive rate of 0.27 percent.
Those two numbers are the kind you can put on a slide. They are also the kind you should read the footnote for.
How the 99.96 percent was measured
For the image evaluation, Pangram took 100 real photographs and created synthetic mirrors of them using 12 different image generators. The prompts were written to closely match each original photo, producing 1,200 synthetic images. On that set, the company reports 99.96 percent accuracy for Image 1.0.
That is a well-designed test for one specific question. Because each real photo has synthetic counterparts of the same scene and subject, the model cannot cheat by learning that beach photos are real and product shots are fake. It has to look at the pixels.
It is not a test of what happens when you point the model at the open internet. Pangram says so directly: the figure describes performance on this targeted pairing of real and synthetic images, rather than a universal accuracy rate for every image encountered online.
The distinction matters more than it sounds. A detector's real-world error profile is dominated by the images nobody thought to include in the test set: screenshots of screenshots, images that went through three messaging apps, old JPEGs, scans of prints, anything with a watermark strip, and the outputs of generators released after the test set was frozen. A clean 100-photo mirror set cannot measure any of that.
The 0.27 percent false-positive rate is the more interesting of the two headline numbers, because false positives are what get people accused. At 0.27 percent, roughly one in 370 real images gets flagged. For a newsroom checking a handful of photos a day, that is a manageable review queue. For a platform scanning a hundred million uploads, it is a quarter of a million falsely accused users, which is why nobody at that scale relies on a single classifier.
The video number is the alarming one
The video benchmark is where this release gets uncomfortable.
Pangram evaluated Flash 1.0 with DF26, a benchmark described in a preprint submitted September 7, 2026. DF26 contains 271 real videos and 2,420 synthetic ones produced by seven modern video models. The content is deliberately narrow and deliberately hard: one person speaking directly to camera, official statements, studio interviews. That is the deepfake format that actually causes damage, because it is the one where a viewer's default assumption is that the person is real.
The benchmark's authors report that both people and state-of-the-art deepfake detectors performed close to random chance on it. They argue for evaluations that explicitly test robustness to distribution shifts from modern generators.
That result should sit in your head for a moment. Near random chance means a coin flip with extra steps. The detectors are well built. They were trained and evaluated on an older generation of fakes, and the newer generators do not leave the same artifacts. A detector that learned to spot warping around the jawline is useless when the jawline does not warp.
Pangram says Flash 1.0 scores 98.7 percent on DF26, which it describes as 29 points above the next-best detector. If that holds up under independent replication, it is a large gap on a benchmark specifically designed to break the previous generation of tools. It is also a company-reported number on a benchmark the company chose, evaluated with its own harness, so the honest position is: promising, unverified.
What the product actually does
The feature list is more informative than the marketing.
Heatmap output localizes synthetic material inside an image. This is the capability that matters for the common real case, which is not a fully generated image but a real photograph with one generated element inserted, or a real image with a generated region painted over. A single yes/no label is useless there. A heatmap tells an editor where to look.
Native video upload runs the frame-by-frame analysis inside the dashboard and identifies which portions of a video come back flagged. For a two-minute clip, that produces a timeline of suspicion rather than a verdict, which is the right output for a human reviewer.
The Chrome extension flags AI-generated material across social media, shopping and rental sites. That last category is the one that costs people money. Rental listing photos are a classic place where a generated image costs somebody real money.
What changed since the July preview
The July 29, 2026 research preview accepted images only at 512 by 512 pixels or larger, and explicitly listed deepfakes and face swaps as cases it could not confidently detect. That was an honest limitation, and it was also a large hole, since face swaps are the most common harmful case.
Version 1.0 claims improved performance on deepfakes, low-resolution images, and frontier generation models. The minimum input is now 256 by 256, with a reduced-confidence disclaimer for anything under 512. The company says it trained the new model to detect transformations such as compression, downscaling, blur and noise, whether introduced by social platforms or by actors trying to evade detection.
That last sentence is doing a lot of work. Most real-world AI images reach you after being recompressed by at least one platform. A detector that only works on pristine PNGs is a lab instrument.
Access has also widened. In the preview, image API access was invitation-only and dashboard video input was listed as future work. Now both exist.
The structural problem nobody has solved
Detection is a classification problem where one side gets to iterate. Every time a generator improves, the artifact distribution shifts, and every detector trained on the old distribution loses accuracy. The defenders have to keep winning. The attackers only have to win once, on the image that matters.
That gap is what provenance standards like C2PA were built to close. A cryptographic signature on a camera capture is a different kind of claim from a statistical guess about pixels. The signature can be stripped, but when it is present, it is not a probability.
The two approaches are complements, not competitors. Provenance answers "do we know where this came from." Detection answers "does this look generated." A newsroom wants both, and a platform wants both, and neither is sufficient.
Pangram's pitch is essentially that detection is not a solved problem and someone should sell it as a product with heatmaps and an API rather than a research demo. That is a reasonable business. The 99.2 percent number will get them meetings. The 98.7 percent on DF26 is what they should be asked about, in detail, by someone with the benchmark harness and a fresh generator.
A practical footnote for anyone using this
If a detector tells you an image is AI-generated, you have a lead. If it tells you an image is real, you have nothing. That asymmetry is built into the problem, and no accuracy figure changes it. The 0.27 percent false-positive rate tells you how often the tool will embarrass you in one direction. Nothing in the announcement tells you how often it misses a good fake, and that number is always worse.
So use it as a triage tool, keep the heatmap, log the model version alongside the verdict, and expect to re-run your evaluation every time a major generator ships. The half-life of a detector's accuracy is measured in months.
Related articles
Chinese text in video always turns to mush. Dreamina handed the typesetting to code
Dreamina launched Motion Studio on November 10. It no longer generates frames; the model writes code that lays out text, shapes and pacing. Chinese titles stop blurring, but whether the content is right is still on you.
A community spreadsheet of taps that turns off the ads inside Chinese Android builds
A no-root tool that switches off the ads Chinese phone makers ship inside the system UI. Its method is a hand-written tap script per phone model, maintained by volunteers every time a ROM moves. It works, and the maintenance is the point.
GPT Image 2.6 held the token price and raised the bill
GPT Image 2.6 kept every token rate from 2.5 and still costs more per image, because more of the work now happens in output tokens. Eleven days into the ramp, the pricing structure says more than the leaderboard does.
Someone rebuilt IDM from the outside and wrote down what they did not touch
A Rust reimplementation of Internet Download Manager claims 5x the speed and 86 languages. The more interesting document is the paragraph explaining how it was built without copying anything.