TwelveLabs Pegasus 1.5 arrives on Amazon Bedrock for long video analysis

From answering questions to finding every matching segment

AWS says video is a large and complex category of unstructured data, yet most of it stays unsearchable and underutilized. Pulling structured insights from it has historically required labor-intensive manual review or fragmented toolchains.

Earlier versions of Pegasus let users point the model at a clip, ask a question and get an answer. Pegasus 1.5 analyzes the entire video, finds every segment that matches user-defined criteria, and returns structured, timestamped metadata. AWS announced the availability on Oct 9, 2026.

The release adds several capabilities:

• Time-based metadata generation: users define segments, events or domain-specific cues, such as editorial segments, sports plays, speaker changes or brand appearances. The model returns matches as JSON with start and end times and the requested fields.
• Multimodal prompting: a segment definition can combine a text description with a visual reference, which makes it possible to locate a specific product, logo or person and what they are doing on screen.
• On-the-fly processing: no ingestion pipeline or pre-indexing step is needed.
• Extended capacity: videos up to two hours long can be analyzed in a single request, up from one hour in Pegasus 1.2.

Pegasus 1.5 keeps the core abilities of Pegasus 1.2: video-to-text generation, granular visual comprehension of objects, on-screen text and numbers, temporal grounding, and multimodal understanding that combines visual, audio and textual information. English is fully supported, with partial support for Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish, Thai and Vietnamese.

How it works on Amazon Bedrock

Developers call the model on the bedrock-runtime endpoint with the InvokeModel API, or with InvokeModelWithResponseStream if they want the response streamed as it is generated. The model ID is twelvelabs.pegasus-1-5-v1:0, the listed AWS Region is US West (Oregon), us-west-2, and inference is on-demand. The AWS example also uses a global Cross Region Inference endpoint, and notes that replacing "global." with "us." targets the US geographic endpoint.

Each request runs in one of two modes, set with the analysisMode field. The default, general, takes a text prompt and returns a text answer, or JSON when a schema is supplied. The time_based_metadata mode takes segment definitions in responseFormat and returns a list of segments with start and end times and the requested fields.

Video can be supplied in two ways. An Amazon S3 location, given as a URI plus the bucket owner account ID, suits long-form video. A base64 string embeds the video in the request body, up to 25 MB, and suits short clips. AWS says videos are analyzed with no pre-indexing, and that data remains within the AWS data boundary.

For cost tracking by application, users can create an application inference profile from the Pegasus 1.5 model, tag it, and call it by its Amazon Resource Name in place of the model ID.

Getting started and what it means for users

AWS lists the prerequisites: an active AWS account with Amazon Bedrock access, Python 3.10 or later with Boto3, IAM permissions for bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream, model access enabled for TwelveLabs Pegasus 1.5 in the Bedrock console, and videos in S3 or short clips to send inline. Authentication can use standard AWS credentials or an Amazon Bedrock API key. For production, AWS recommends a short-term key because it expires automatically.

In general mode, temperature ranges from 0 to 1 with a default of 0.2, and maxOutputTokens ranges from 512 to 98,304. A stopReason of length means the answer hit the token limit before finishing, so developers should check it before relying on the output. Setting includeMetering to true adds outputTokenCount and inputVideoSecond, the analyzed video's duration in seconds, to non-streaming responses.

For teams that handle large video libraries, such as media, sports or marketing workflows, the change is that a single request can return structured, timestamped results that code can parse directly, without building an indexing pipeline first. The source does not give pricing.