TwelveLabs launches Pegasus 1.5 for time-based metadata

Hot take: this is a real product shift, not a cosmetic model bump. TwelveLabs is trying to make video behave more like a database table than a blob of media.

• –The core change is boundary detection plus structured extraction across an entire video, which is much more operationally useful than clip-level Q&A.
• –The schema-first design is the right move for enterprise workflows where “what matters” differs by domain, whether that is sports, media, or brand monitoring.
• –The model is clearly aimed at downstream systems, not demos: the emphasis on valid JSON, timestamps, and non-overlapping segments matters more than fluent prose.
• –The main moat is not just model quality, but the evaluation and training stack built around temporal metrics and verifiable rewards.
• –If Pegasus 1.5 holds up in practice, it could reduce a lot of custom preprocessing and annotation glue code for teams working with large video libraries.