What did Pangram release on 4 November 2026?
Pangram released Pangram Image 1.0 and Pangram Image Flash 1.0 on 4 November 2026. The first model examines still images. The second is a smaller, faster model for video. Both are now available through Pangram's web dashboard, its API and its Chrome extension.
The release moves image detection out of the research preview Pangram opened on 29 July 2026. That preview accepted images, returned a score and offered a heatmap. The heatmap shaded the regions the model considered synthetic. The new version keeps that feature and applies it to mixed media, where an AI object or edited region sits inside an otherwise ordinary photograph.
Flash 1.0 handles video by analysing frames. A user can upload a clip, and the dashboard reports which parts of the footage come back as AI. The updated API accepts both images and video, so a platform can call the detector inside its own review pipeline. That shift from a consumer checker to an integration layer is the practical change in this release.
Pangram also lowered the minimum image size from 512 by 512 pixels to 256 by 256 pixels. Images below 512 pixels still carry a reduced-confidence warning. The company says the new model is more accurate at low resolution, which is why the floor moved. The warning remains because a small image contains less evidence.
The company's background helps explain the distribution push. Stanford machine-learning graduates Max Spero and Bradley Emi founded Pangram roughly two years before July 2026. TechCrunch reported a $9 million round led by Menlo Ventures that month, with Haystack, ScOp, Script Capital and Cadenza participating. At the time, the web product cost $20 per month.
Pangram's earlier work was text detection. The image model extends the same commercial idea to pictures. The pitch is that a reader, platform or employer can ask a simple question: was this made by a machine?
How accurate is Pangram Image 1.0?
The headline number is 99.96%. Pangram built the evaluation set by taking 100 real photographs and using 12 frontier image generators to create a synthetic mirror of each one. The prompts matched the original scene closely. The result was 1,200 synthetic images.
The model classified that set, and Pangram reports 99.96% accuracy. The company describes the design as a way to prevent shortcuts. Each real image and its synthetic twin share a subject and scene, so a detector cannot score well by recognising a genre. It has to learn the difference between a camera and a generator.
Unite.AI, citing Pangram's launch material, reports a second figure: 99.2% of AI-generated images detected at a 0.27% false-positive rate. The false-positive rate is the number that matters to an innocent user. At 0.27%, roughly one in 370 human images would be flagged as synthetic.
The July preview offers a longer track record. Pangram said the preview reached 99.5% accuracy against five commercial detectors, with the next closest at 98%. It reported above 99% AUROC on NTIRE 2026, Synthbuster and Mirage. It also reported a 0.16% false-positive rate on ReLAION, a set of pre-2022 internet images, or about one in 625.
Pangram's comparison, as reported by Digital Trends, placed TruthScan at 96.33% on images. AI or Not followed at 91.38% and Sightengine at 88.33%. Those figures describe performance on a set Pangram designed. They show a gap between products, not a guarantee about the next image a user uploads.
AUROC and accuracy measure different things. AUROC describes how well a model ranks synthetic images above real ones across thresholds. Accuracy describes decisions at one threshold. A product can look strong on both and still produce false alarms when the input distribution changes.
Why is AI video detection harder than image detection?
Video is a harder problem. A clip carries motion, changing compression and hundreds of frames. A detector has to aggregate evidence across time without letting one bad frame decide the result. Pangram's answer is a lighter model that examines frames separately and reports the parts of a video that look synthetic.
For its industry comparison, Pangram started with 129 real social-media videos. It used 14 frontier video generators to create synthetic mirrors, sampling 100 clips per generator from the real set. The total was 1,529 videos. Each generator received the first real frame and a vision-language description of the action.
In that comparison, Digital Trends reports 98.57% accuracy for Flash on AI videos. It puts TruthScan at 84.2% and Sightengine at 81.8%. The same article reports Pangram's claim of 98.7% on DF26, described as a 29-point lead over the next-best detector.
DF26 is the more interesting test because it did not come from Pangram. The benchmark's preprint appeared on arXiv on 7 September 2026. It contains 271 real videos and 2,420 synthetic clips from seven modern video models. The material covers direct-to-camera recordings, official statements and studio interviews.
The authors' finding is stark. Human viewers performed close to random chance when judging whether a public-speaking video was real. State-of-the-art deepfake detectors were not much better. That result explains why a strong benchmark score attracts attention. It also explains why the benchmark exists: older tests were too easy.
A 98.7% score does not mean the model catches 98.7% of every fake in circulation. It means it caught that share of the DF26 sample at a chosen threshold. The paper's authors argue for tests that measure robustness to new generators, not just accuracy on known ones.
What do the benchmarks actually prove?
Accuracy hides class balance. Pangram's image set contains 1,200 synthetic images and 100 originals. If the score is computed across all 1,300, a detector that called every image AI would already reach about 92.3%. The 99.96% figure is better than that baseline, but the raw percentage exaggerates the gap.
DF26 has a similar skew. With 2,420 synthetic clips and 271 real ones, an always-AI classifier would score about 89.9%. Flash's 98.7% still beats that floor, and the 29-point lead over competitors is meaningful. The absolute number is less informative than the comparison.
The missing metric is the false-positive rate on real video. Pangram publishes a false-positive figure for images, through its launch material and the preview. Its video post reports accuracy and a lead over rivals, but not a rate for real clips. A platform needs that number before it labels or removes content at scale.
The comparisons are also vendor-run. Pangram chose the generators, the prompts and the competitor settings. DF26 is independent, but Pangram chose to be measured on it. None of the launch results has passed peer review. That does not make them wrong. It makes them provisional.
The preview was candid about scope. It could not confidently detect deepfakes or face swaps, because it focused on consumer image generators. It blocked NSFW content. Its API was invitation-only, and its Chrome extension did not yet include images. Unite.AI reports that Pangram claims improved performance on deepfakes and low-resolution pictures. The launch post does not give a separate score for face swaps.
Independent hands-on testing adds texture. TechCrunch's reporter found the preview's image detection impressive. The heatmap correctly lit up an AI image placed inside a real photograph. In one case, though, the same kind of photo was labelled human. A single miss is not a verdict on the model, but it shows why scores need calibration.
A benchmark is a controlled sample. The open internet is not. Images arrive cropped, screenshotted, recompressed and mixed with text. A model trained to resist those changes is better equipped than one that is not. It still faces a distribution it did not choose.
Who is affected by better AI-image detection?
The most immediate users are platforms and marketplaces. Pangram's extension is designed to flag AI content while someone scrolls social media, shopping pages or rental listings. In each setting, a fake image can extract money or trust. A rental listing with a generated interior is a fraud risk, not an aesthetic one.
Newsrooms, insurers and courts face a higher-stakes version. A detector score is one piece of evidence. It is not a chain of custody. An editor still needs to know who supplied the image, when it appeared and whether an original file exists.
Regulation is adding pressure. Article 50 of the EU AI Act requires disclosure when image, audio or video content is a deepfake. The rule carves out evidently artistic, satirical or fictional work, where the obligation is limited to disclosing that the content was generated. The remaining provisions of the Act, including this one, began applying on 2 August 2026.
That timing makes detection commercially attractive. A platform or publisher can use a detector to screen material before it publishes. The tool does not confer compliance by itself. Regulators will ask what a company did with a positive result, not whether it owned a detector.
False positives carry real costs. A photographer or illustrator accused of using AI can lose work or audience. Pangram addressed this in the preview, saying it classified all 2,000 images in a WikiArt subset as human. That is a strong result on a curated art collection. It is also a company-run test.
Access shapes who benefits. TechCrunch reported a $20 monthly subscription and API customers including Quora, schools and universities, publishers and agents, and recruiters. Organizations can integrate detection. Individual readers mostly get a browser extension and a score.
What are the limits of AI detection?
Detection works by learning statistical traces left by known generators. A new model, a new sampling method or a heavy edit can change those traces. Pangram says it trains against compression, blur, noise and downscaling. That covers common transformations, not every future one.
Small images remain weak. The 256-pixel floor is lower than before, but the company still attaches a reduced-confidence warning below 512 pixels. Screenshots, avatars and cropped evidence often land in that range. A confident label on a low-resolution image would be misleading.
Composite images are another edge case. Pangram says AI-edited images often score as fully AI because modern generators rebuild the whole frame. The heatmap helps a human see where the model reacted. It does not turn a percentage into a legal finding.
Provenance offers a different route. Watermarks and signed metadata can show which camera or tool produced a file. Pangram argues that watermarks are easily stripped and that not every provider adds them. Its pitch is that pixel-level detection still works when the signature is gone.
The methods are complementary. A provenance chain is strong evidence when it survives intact. A detector is a fallback when the chain breaks. Neither replaces human judgement in a dismissal, a prosecution or a front-page decision.
The honest summary is that detection raises the cost of deception. It does not end it. A high score shifts the odds. It does not establish authenticity.
What should happen next?
The next step is independent evaluation on balanced sets. Researchers should report true-positive and false-positive rates separately. They should test generators the vendor did not select. DF26 is a useful model because its authors published the construction method and the class counts.
Platforms should disclose how they use scores. A warning label, a downrank and an account suspension carry different error costs. A 0.27% false-positive rate may be acceptable for a caution. It is harder to justify as the sole basis for a ban.
Buyers should ask for the test set, not just the headline. Pangram's 99.96% and 98.7% are traceable to described evaluations. That is better than many vendor claims. The descriptions still refer to curated mirrors, not the open web.
The arms race will continue. Each new generator supplies detectors with fresh training data and gives bad actors a new way to probe evasion. The useful question is how quickly a detector degrades when the generator changes.
For readers, the practical rule is modest. Treat a detector result as one signal. Check the source. Look for provenance when it exists. A high score is not a certificate, and a low score is not proof of fraud.
Pangram's release is a real advance in access. Detection now sits in a browser, an API and a dashboard. The benchmark claims deserve scrutiny, but the underlying problem is not going away. The tools will improve, and the tests should improve with them.