Model releases, new detectors and API changes, newest first.
Videos encoded with AV1, common on YouTube and Instagram, are now decoded and checked instead of returning an error. Read more
Logos, a founder photo and numbers you can quote, each linked to its method. Read more
Evaluation code for text, image and video detectors, and SHA-256 hashes of the hidden test split so it cannot change unnoticed. Read more
What each product stores, which providers process data, and what enterprise customers get. Read more
Pages for investigations, trust and safety, social media, newsrooms, KYC, marketplaces and recruiting. Read more
The image detector now flags an image from a score of 0.8 instead of 0.5, which cuts false positives on real web and phone photos.
POST /api/v1/detect/video scores whole clips across sampled frames, or a single frame. Upload up to 4MB or pass a URL for files up to 50MB. Read more
A clip costs 300 credits per frame the model checks, 4,800 to 15,000 credits depending on length. The detector shows the price before it runs.
19 image generators and 12 video models, never used in training. 60% of the test set is public on Hugging Face. Read more
Free API credits for researchers who audit our text, image and video detectors. Read more
$299 a month with 3,000,000 API and MCP credits. Existing Unlimited API subscribers keep their plan.
A new detector for images, single frames and video clips, trained on the newest generators and on content that went through messenger and social-media compression.
Images and audio are also checked against OpenAI’s content provenance signals. A confirmed signal marks the result as verified.
Each API key can make 120 requests per minute. Above that the API returns 429 with a Retry-After header.
At a score of 90, tropa-3 flags 0.1% of held-out human texts, down from 3.02% for tropa-2. Read more
2,572 texts including six humanizers, with every text and verdict public. Read more
Checks can be exported with a signed badge that links to a public record of the result.
AI agents such as Claude can call the detector directly, with OAuth sign-in. Read more
Sentence-level scores, a calibrated 0 to 100 score and a hint when text looks humanized. Read more
C2PA manifests are parsed and their signatures verified. A valid AI declaration is reported as certainty, not a score.
A full review of theses and long papers, in any language, with a PDF report. Read more