news

Pangram's AI video detector launches with two accuracy figures and two different runner-up scores

The launch blog claims 98.7% on the DF26 benchmark and a 29-point lead, while the company's own X thread says 98.57% against a next-best competitor at 84.2%.

By Daniel Ochoa · 6 November 2026

Pangram, known until now for detecting AI-written text, released two vision models on November 4, 2026: Pangram Image 1.0 for stills and a lighter Image Flash 1.0 that runs video frame by frame. The company calls the pair the state of the art in both AI image and AI video detection. The problem for anyone trying to use that claim is that Pangram's blog and Pangram's X thread, published the same day, do not agree on what the video model scored or on how far ahead it finished.

98.7% with a 29-point lead, or 98.57% against 84.2%

The launch post puts Image Flash 1.0 at 98.7% on DF26 and describes that as a 29-point leap over the next best detector, which works out to roughly 69.7% for the runner-up. The company's X thread states 98.57% accuracy on AI videos versus 84.2% by the next best competitor. Neither channel says whether the two sets of numbers come from the same evaluation. If they do, the headline gap being claimed differs by about 15 points depending on which Pangram page you read. The naming is inconsistent too: the blog calls the model Pangram Image Flash 1.0 and the X thread calls it Pangram Image 1.0 Flash.

DF26 covers talking heads, not video in general

DF26, the one external benchmark in the announcement, comes from a preprint submitted in September 2026 and contains 271 real and 2,420 synthetic clips, 2,691 in total, produced by seven text-to-video and image-to-video models. Its scope is one person speaking to camera: direct-to-camera recordings, official statements, studio interviews. Each synthetic clip is generated from a prompt derived from the first frame of a real one, holding identity and scene constant so the generator is the variable. That design makes it a clean test of a narrow case, and it is not a measurement of detection across landscapes, crowds, product shots or animation. The benchmark's authors reported that humans scored barely above chance on this material and that existing detectors did only slightly better.

Direct-address speech is the exact footage class in question, and it is one our own prompt set generates, as in this Sora 2 clip:

The image number comes from a test Pangram built

For stills, Pangram reports 99.96% accuracy on a set it constructed by mirroring 100 real photographs across 12 frontier image generators, with prompts written to match each original, yielding 1,200 synthetic images. Holding the subject matter constant is a sound way to stop a detector from winning on content cues rather than generation artifacts, but the set was designed, generated and scored in-house. Separately the company reports identifying 99.2% of AI images at a 0.27% false positive rate. We found no third-party evaluation of either model.

What actually shipped

Image 1.0 now runs inside Pangram's Chrome extension, flagging suspected AI material as users browse, and it accepts images down to 256 by 256 pixels, down from a 512 by 512 floor, with a reduced-confidence warning below 512. Users can upload video on the website for frame-by-frame analysis, and the API was updated to cover both images and video. No API price appears in the announcement. A heatmap output localizes which regions of an image or frame the model considers synthetic, which is a more useful artifact than a binary confidence score and also a harder one to score against a benchmark.

The context here is a field where the published numbers keep coming from the vendor. A Kling-affiliated detector recently reported 89.06 balanced accuracy with a far lower score on locating artifacts, and in October OpenAI and Meta told a US evidence-rules panel that no method reliably detects AI video. A 98.7% on talking heads, reported twice with different digits, does not settle that.

Sources: Pangram, Engadget, Digital Trends.

Daniel Ochoa: Covers model launches, shutdowns and pricing changes as they happen. Reads deprecation notices for a living so you do not have to.