TwelveLabs Pegasus 1.6 reads first-person video at $1.75 an hour

TwelveLabs released Pegasus 1.6 on October 6, a video-understanding model that turns first-person footage into labelled, timestamped training data for robots.

The three rates match the version it replaces. TwelveLabs' pricing page lists Pegasus 1.6 at $1.75 per hour of input video, $3 per million image-input tokens and $7.50 per million output tokens. Those three rates are the same list TwelveLabs published for Pegasus 1.5, the model Pegasus 1.6 replaces.

What Pegasus 1.6 adds

The main change is egocentric, or first-person, video. TwelveLabs' model documentation says Pegasus 1.6 analyzes footage from wearable cameras and teleoperated systems, and names action labeling, quality control and next-action prediction as new use cases. The release post says earlier versions did not support that category well, and the release notes date the model to October 6, 2026.

The update also brings native image analysis of one to twenty stills per request, stronger entity recognition, timestamped in-segment events and segmentation without a token limit. The model keeps a context window of 261,120 tokens and accepts video up to two hours long, or four hours when only part of it is analyzed, at a file size up to 10 GB.

What 1,000 hours costs

TwelveLabs bills Pegasus by video duration and by output tokens, not by input tokens. At the listed video rate, a single pass over 1,000 hours costs $1,750 before output and other charges, a figure the maker's own numbers give directly: 1,000 times $1.75. A hundred hours costs $175.

Segment requests carry a multiplier on top. TwelveLabs' pricing FAQ says each segment definition is billed separately, so one hour of video split by four definitions bills as four hours, or $7 at the listed rate. A team asking for several action categories pays for each of them.

Pegasus describes and organises video; it does not drive a robot. VentureBeat notes the model supplies descriptions and structure, and robotics teams still have to connect those observations to movement, sensing and control. TwelveLabs co-founder and CEO Jae Lee told the outlet that human video could complement teleoperated demonstrations, and said in the company's release that most knowledge of doing physical work "has never been captured in a form a machine can learn from."

The company arrives from a different market. Lee told VentureBeat its early customers came from media, entertainment, sports and advertising, where the job was searching and describing large video collections. The robotics pitch applies those same tools to physical work, and TwelveLabs' physical AI page lists five workflows: labelling actions, writing captions, screening footage for quality, finding unusual events and duplicates, and flagging sensitive material.

For teams weighing the pipeline, the deciding number is cost per usable training example, not the headline rate. Pegasus 1.6 runs on the same API as the version before it, so the price list above is what a task costs today, absent an enterprise contract, which TwelveLabs prices on request.

Pegasus 1.6 works with existing Pegasus code; TwelveLabs' release notes say to run pip install --upgrade twelvelabs, or npm install twelvelabs-js@latest for Node.js, and new accounts start with up to 10 hours of free indexing.