TwelveLabs Pegasus 1.5 launches on Amazon Bedrock with two-hour video segment retrieval

Pegasus 1.5 is TwelveLabs' video language model, invoked via bedrock-runtime's InvokeModel or InvokeModelWithResponseStream with model ID twelvelabs.pegasus-1-5-v1:0, currently offered only in US West (Oregon) as on-demand inference. It has two analysis modes: general mode takes an inputPrompt and returns text or JSON Schema-constrained structured output; time_based_metadata mode takes segment definitions (id, description, fields) and returns start/end times and the requested fields for each matching segment. Videos can be read directly from an S3 URI or submitted inline as a base64 string under 25 MB, with no pre-indexing. Compared with Pegasus 1.2, the new version analyzes up to two hours of video per request, adds video segmentation, multimodal prompts with reference images, and instant processing, and supports locating specific products, logos, or people using text plus reference images in segment definitions. English is fully supported, with partial support for Chinese, Japanese, French and a dozen other languages. For teams building video retrieval pipelines on AWS, the key change is dropping the ingestion and pre-indexing step, keeping data within the AWS boundary, and using includeMetering to return outputTokenCount and inputVideoSecond for per-video-second cost accounting; teams that tag costs via application inference profiles can switch to ARN invocation.