What happened
On November 4, 2026, Pangram announced Pangram Image 1.0 and Pangram Image Flash 1.0. The company says Image 1.0 is its flagship detector for AI-generated images and can return a heatmap to localize where AI-generated content appears within an image. It also says the model is designed to handle edits, compression and low resolutions. Pangram describes Image Flash 1.0 as a smaller, faster model for video detection, with frame-by-frame analysis. The supplied announcement does not give pricing or usage limits for either model.
Pangram reports 99.96% accuracy for Image 1.0 on an evaluation set it constructed. The set contains 100 real photographs and synthetic mirrors made with 12 frontier image generators, for 1,200 synthetic images in total. The company says the prompts for those synthetic images closely matched the original photographs. It also says it compared its image model with commercial and academic detectors and evaluated against several academic benchmarks, naming X-AIGD, Mirage-Test, AIGenImages2026, Synthbuster + RAISE comparison, and Community Forensics. The announcement does not provide a full results table for those comparisons in the supplied material.
The release changes the stated minimum supported image size: Pangram says its initial preview limited input to 512×512 pixels, but Image 1.0 lowers that floor to 256×256. Results for images below 512 pixels in either dimension carry a reduced-confidence disclaimer. The company says it trains for transformations such as blur, noise and other artifacts that may occur in social-media processing or deliberate image alteration, and says it found the model robust to those edits. It does not quantify performance for each transformation in the announcement.
For video, Pangram cites the academic study DF26, saying humans were barely better than random at distinguishing real videos from AI videos of people talking, while modern detectors also struggled. Pangram reports that Image Flash 1.0 scored 98.7% on DF26, 29 percentage points above the next-best detector. In a separate company-run comparison, it used 129 real social-media videos and generated AI mirrors with 14 video generators, 100 videos per generator, for 1,529 videos overall. Each synthetic video began from the first frame of a real video and used a description generated by a vision-language model. Pangram says pairing each real video with a synthetic counterpart was intended to reduce bias toward labeling certain subjects or scenes as AI.
The announcement adds native video uploads, with Flash analyzing footage frame by frame, and says the API now supports image and video detection. Image 1.0 is also integrated into Pangram’s Chrome extension, which the company describes as a way to flag AI-generated content while browsing social media, shopping, rental listings and other sites. It says Flash achieves state-of-the-art results on several academic video benchmarks, naming AIGVDBench, DF26 and Ivy-Fake. Access is described through the extension and API, but the supplied source gives no API price, quota, service-level commitment or detailed access terms.
What the archive adds
The archive has covered AI-generated image systems, including Fable 5.1’s image-embedded messages and Linum’s text-to-image model. Those are adjacent topics, not earlier coverage of Pangram’s detector. This is a new product event and opens a distinct thread on detecting generated media.
Trend signal
No supplied SITE_PATTERNS thesis directly concerns AI-content detection. The pattern about generative media becoming application infrastructure is specifically about composing generation as controllable infrastructure, so it does not fit this detection release closely enough to claim support. The product’s API is relevant context, but not evidence for that pattern’s thesis.
Why it matters
For teams that publish, moderate or review images and videos, an API and browser extension make detection easier to trial in existing workflows than a standalone research result. Heatmaps may also help reviewers inspect why a detector flagged a particular image, rather than treating a single label as self-explanatory. Lowering the stated input floor to 256×256 could make the tool applicable to smaller assets, though the confidence warning below 512×512 matters for decisions where false positives carry a cost.
The video results are promising enough to motivate a controlled evaluation, not enough to justify treating the detector as ground truth. A team can compare Pangram against its own mix of authentic and generated content, including edits and platform compression, and decide where human review remains necessary. The release does not report how performance translates to a production workload, so buyers should assess latency, cost, throughput, and error rates before building a dependency around the API.
Limits and uncertainty
The announcement is Pangram’s account of its own release and evaluations. Its reported 99.96% image accuracy and 98.7% DF26 video score do not establish how the models will perform on every generator, camera, platform, language-independent workflow or adversarial edit. The supplied material does not include confidence intervals, a complete comparison table, or enough detail to independently reproduce the company’s evaluation. It also does not state price, API quotas, latency, data-retention terms, or availability by plan. Lower-resolution image results explicitly carry a reduced-confidence disclaimer. Detection scores should not be read as proof of origin for an individual file.
Skill to build now
Build a small, labeled evaluation set from the images or videos your team actually handles. Include authentic media, generated examples from relevant tools, resized or recompressed files, and edited content. Run candidate detectors on the same set, record false positives and misses separately, and check whether a heatmap helps reviewers understand the result. Keep the set separate from any material used to tune a workflow, and define which decisions require human review. This turns a vendor benchmark claim into evidence about your own operating context.
Reader takeaway
Pangram’s release makes image and video detection available through a browser extension and API, with heatmaps for images and frame-level video analysis. The reported benchmark results are a reason to test the product, not a substitute for testing it: performance, confidence and acceptable error rates depend on the media your team actually needs to review.