On November 4, 2026, Pangram shipped Pangram Image 1.0 and a faster video-focused model, Pangram Image Flash 1.0 — reporting 99.96% image accuracy, AI-localizing heatmaps, and frame-by-frame video analysis.
2026-11-06 · by Moe Ameen
On November 4, 2026, Pangram released Pangram Image 1.0, which it calls the state of the art in both AI image and AI video detection. It's the production version of the AI image detector the New York company — founded by Stanford AI graduates Max Spero and Bradley Emi — first shipped as a research preview in mid-2026, the same stretch it raised a $9 million round led by Menlo Ventures. Alongside the flagship image model, Pangram released Pangram Image Flash 1.0, a smaller, faster model built for video.
The headline claim is accuracy. Pangram reports 99.96% accuracy for Image 1.0 on its own evaluation set — 1,200 synthetic images it built by mirroring 100 real photos with 12 frontier image generators — and says the model holds up across academic benchmarks including Synthbuster, Mirage, and Community Forensics. For video, it says Flash 1.0 scores 98.7% on the DF26 benchmark, which it describes as a 29-point lead over the next-best detector. These are company-reported figures on a hard, fast-moving problem.
Two practical changes stand out. Image 1.0 outputs a heatmap that localizes where AI content appears in a picture, so it can flag an AI element composited into an otherwise real photo instead of returning a single yes/no. And the minimum input size drops from 512×512 to 256×256 pixels — results under 512 carry a reduced-confidence note — which matters for the compressed, re-saved images that actually circulate on feeds. Pangram also says the model is trained for robustness to compression, edits, blur, and noise.
You reach it through Pangram's web app (image detection plus frame-by-frame video upload), a browser extension that flags AI content as you browse, and an API that covers both image and video. The announcement links to a general pricing page but does not spell out the limits of the visual-detection tiers.
The signal here isn't "hide your AI better." It's that visual provenance is becoming a reader-facing feature the way text detection already did — and the durable response is the same one that worked for writing: produce genuinely on-brand visual content, disclose how you make it, and stop leaning on anonymous, generator-default output that reads as AI at a glance. That's what [Kompozy](/) produces. Where Pangram Image 1.0 scores a picture or a clip after the fact, Kompozy generates the visual in a consistent identity to begin with: a Gemini face-locked Persona Photo, a HeyGen avatar Persona Short, a brand-exact [HyperFrames](/glossary/hyperframes) carousel, or a [Clipped Short](/glossary/clipped-short) cut from long-form footage — each held to one [Persona Brief](/glossary/persona-brief) so the output is recognizably yours, not a stock render.
Act on the news as a workflow, not a worry. Generate your visual set in Kompozy, keep a clear "how this was made" disclosure, and use Pangram as an optional proofing gate on the pieces where it matters — reading the heatmap to see which element a viewer's detector would light up, then regenerating that asset instead of guessing. Kompozy reframes each visual to 9:16, 1:1, and 16:9 and schedules it across eight social platforms plus blog and email behind a per-post review pipeline with [Autopilot](/glossary/autopilot). Pangram tells you how a visual reads; Kompozy makes the visual — and the content around it.
It is Pangram's production AI image detector, released November 4, 2026, which the company calls state-of-the-art for both AI image and video detection. It returns a likelihood that a picture is AI-generated and outputs a heatmap showing where AI content appears, and it ships alongside Pangram Image Flash 1.0, a faster model built for video.
Pangram reports 99.96% accuracy on its own evaluation set of 1,200 synthetic images and says Flash 1.0 scores 98.7% on the DF26 video benchmark. Those are company-reported numbers on a hard, fast-moving problem, so treat any single result as an estimate rather than proof.
Yes — Pangram Image Flash 1.0 analyzes uploaded video frame by frame so you can see which parts are flagged, and Pangram reports state-of-the-art results on video benchmarks including DF26. Image 1.0 handles stills, with a minimum input size lowered to 256×256 pixels.
Not by laundering output. The durable move is to produce genuinely on-brand visuals and disclose your process — an engine like Kompozy generates face-locked Persona Photos, avatar video, and brand-exact carousels, then publishes them across platforms, with a detector used only as an optional proofing step.