Pegasus 1.5

Pegasus 1.5, released on 20 April 2026, is the text-generating half of the TwelveLabs platform: where Marengo turns video into vectors for search, Pegasus turns it into sentences. It reads picture, speech and on-screen text together and answers questions about the recording — what happens, when it happens, what is written on a sign in the third minute. The headline addition in this version is video segmentation, and it is more useful than the name suggests. Rather than asking for a summary, the caller defines the kinds of segments to look for — editorial narratives, sports plays, speaker changes, brand appearances — along with the fields each segment should carry, and gets structured JSON back. That turns the model from a describing tool into an indexing tool: raw footage in, timestamped records out. Time ranges can be set per definition, so one pass can look for different things in different parts of a recording. Two other changes matter in practice. Pegasus 1.5 analyses a video straight from a URL, an asset or a base64 string, with no pre-indexing step — version 1.2 required the file to be indexed first. And prompts can include reference images, so a question can point at a particular person, object or logo instead of describing it in words. The limits are stated precisely by the vendor. A request shares a context window of 261,120 tokens between the video, its transcript, the prompt, any reference images, the schema and the answer, with responses up to 98,304 tokens. Video may run to two hours, or four hours if only a portion is analysed, up to 2 GB. Language support is uneven and admitted as such: English is fully supported, twelve further languages — including Chinese, Japanese, Korean, Arabic and Spanish — only partially. The predecessor is gone. TwelveLabs removed Pegasus 1.2 on 18 August 2026; the platform now rejects requests naming it and falls back to 1.5 by default. Weights are not published. Billing for video analysis is $1.75 per hour of input video plus $7.50 per million output tokens, with a free tier of ten hours of indexing. The model is also offered through Amazon Bedrock.

▸Videos

No videos yet.

▸Reviews