As AI-made photos and videos become indistinguishable from real ones, the numbers on the detection side grow just as fast. The question is who wrote the test behind a detector's "99%." Pangram's Image 1.0 and Image Flash 1.0, released on November 4, face that question head-on. Before the scorecard, read this first: one top score comes from a test the company designed itself; another comes from an academic benchmark but was announced by the company; and the per-event scores for the remaining academic benchmarks still live only inside chart images. This article separates the per-unit prices the official price list settles, the design of the tests the company published, and the ground nobody outside the company has re-measured yet.
The short version
Instead of a one-line verdict, Image 1.0 returns a heatmap that localizes where AI appears in an image, and Flash 1.0 analyzes video frame by frame (official blog, checked directly)
Prices are per-unit on the official price list — $0.25 per image for Image 1.0, $0.05 per image for Flash, $0.05 per second for video (re-checked on Nov 9, 2026 before publication)
Read every score by who wrote its test — 99.96% is from the company's own evaluation set, 98.7% is a company-announced result on the academic DF26 benchmark, and per-event academic scores exist only in charts
The point of Image 1.0 is not its accuracy figure but the shape of its output. Where earlier detectors returned a one-line "this is AI," 1.0 returns a heatmap showing where AI appears inside the image. When only part of a photo is the problem — a marketplace photo with one object erased, a listing photo with only its background swapped — a bare verdict gives a human no next action. The separately trained robustness to edits, compression, and downscaling belongs to the same logic: images on social platforms are compressed, resized copies, and a detector that only works on clean originals loses its power in the field.
Video is Flash 1.0's job. The smaller, faster model analyzes video frame by frame and reports which segments read as AI, and the dashboard now takes native video uploads. The input floor dropped too: the preview's 512×512 minimum fell to 256×256. Below 512, though, results carry a reduced-confidence disclaimer. Lowering the floor and being trustworthy below it are two different sentences.
Prices sit on the official price list as per-unit rates, separate from subscriptions.
| Item (Pangram official price list, checked directly) | Rate | Notes |
|---|---|---|
| Image 1.0 (image detection) | $0.25 per image | Flagship, heatmap included |
| Image Flash 1.0 (image detection) | $0.05 per image | Smaller, faster model |
| Video detection | $0.05 per second | A 60-second clip converts to $3.00 |
| Text detection (Pangram 4.1) | $0.05 per 100 words | Bulk API at 20% off |
On subscriptions, Individual at $20 a month (after a 7-day free start) includes $10 of monthly API credits, and Professional at $65 a month includes $200. On unit rates alone, scanning 1,000 photos costs $250 on Image 1.0 and $50 on Flash. Video bills by the second, so length is price: a 10-second clip is $0.50, a 10-minute video is $30. Bulk-screening short social clips and scanning long videos end to end are entirely different budgets on this price list.
Set against Sightengine, a same-tier service, the first thing that differs is the grammar of billing. What Sightengine's official pricing page shows directly is a feature matrix: AI image detection is open from the free tier, AI video detection from Starter up, video files cap at 50MB on Starter and 500MB on Pro, frame analysis adjusts between 0.5 and 2 fps, and throughput runs at 1 request per second on free and Starter and 10 per second on Pro. The amounts on that page could not be confirmed in text form even in the pre-publication re-check on Nov 9, 2026, so this article prints no Sightengine prices. Pangram counts in images and seconds while Sightengine counts in operations, so until someone settles how many operations one scan converts to, declaring a winner on price alone is a category error.
The performance numbers split like this. The 99.96% is accuracy on an evaluation set the company designed itself: 100 real photos mirrored, with closely matched prompts, by 12 frontier generators into 1,200 synthetic images. The paired design does try to be fair — mirroring blocks shortcuts based on scene or subject bias. Still, the test's author and its taker are the same party. The 99.2% identification rate and 0.27% false-positive rate the company is said to have announced at launch on X remain, in this pass, figures carried by reporting rather than checked against the original post. The video figure of 98.7% is different in kind: DF26 is an academic benchmark (a preprint), not a Pangram test, and its premise is that both humans and existing detectors performed barely above chance on it. But the score itself, and the claimed 29-point lead over the runner-up, are still the company's announcement. The per-event scores behind the "state of the art" claims on academic benchmarks (X-AIGD, AIGenImages2026, AIGVDBench, Ivy-Fake, and others) exist only in the blog's chart images, so this article confirms no numbers for them.
There are three doors in. First, the Chrome extension, which flags AI-generated content while you scroll social feeds and shopping or rental listings — the exact spot where, in Korea, it maps onto screening fake photos in secondhand listings and stays before purchase. Second, native video upload in the dashboard: Flash returns the segments it detects as AI, so a reviewer can narrow in instead of rewatching a whole short-form clip. Third, the API, whose documented uses are review queues, content pipelines, and coursework checks in education, with a zero-data-retention option and SOC 2 on a separate enterprise track. The design treats detection as a signal for a human queue, not as a final verdict.
Outside validation so far belongs to the text detector. A review's account of the 2025 NBER working paper 34223 (Jabarian and Imas) rates Pangram's text detector ahead of Originality.ai, GPTZero, and a RoBERTa classifier — but that is not a score for the image model. No independent lab has yet re-measured Image or Flash 1.0 on its own data, and no named platform has published the scale at which it filters with this model. The case section is thin across this whole product category: detectors are tools whose adoption customers prefer not to advertise, so the weight of proof lands, in the end, on re-measuring with your own data.
First, who designed the test. The 99.96% starts from 100 real photos in a set the company built; however fair the mirroring, the choice of 12 generators, the prompts, and the 100 photos were all the company's, and the compared detectors' names and scores live only in a chart. That an external test track exists, and that its scores were published only as charts, are two facts to hold together. Until scores arrive in re-measurable form, the accurate summary is "very high on the company's test; external per-event scores unconfirmed."
Second, the false-positive rate. In production, the number that hurts is not accuracy but false positives: real photos wrongly flagged mean sellers, students, and journalists flagged by mistake. The advertised 0.27% could not be checked against the original X post in this pass, and which evaluation set it is based on is unconfirmed. Anyone evaluating adoption should first run several hundred of their own real photos and count the flags — before trusting any announced figure.
Third, the unit trap. The $0.25 image rate and $0.05 per-second video rate are settled, but a competitor's operation unit does not convert without a table, and inside Pangram itself video cost scales with length. The detailed allowances where subscriptions and per-unit billing overlap, such as included image scans per plan, also remain unconfirmed in this pass. Budget from "how many images and seconds per month," not from the subscription sticker.
The conclusion splits cleanly. Confirmed: an image detector that localizes AI content with a heatmap and a small model that reads video frame by frame have formally launched; their rates — $0.25 and $0.05 per image, $0.05 per second of video — are on the official price list; and the input floor fell to 256×256 with a confidence disclaimer below 512. Not yet confirmed: real-world accuracy and false-positive rates outside the company's own test design, per-event academic benchmark scores, and quality on Korean content and Korean platform images. The test is in your own data: build an evaluation set mixing your real photos and videos with synthetics from the generators you actually use, and count false positives before accuracy. A detector's price is set not by someone else's 99.96%, but by its false-positive rate on your data.
Puzz, in one line
Judge this detector not by its scorecard but by who wrote its test. Pangram Image 1.0 is a tool whose price must be set by its false-positive rate on your own data, ribbit.