AWS adds TwelveLabs' Pegasus 1.5 video model to Bedrock
Timestamped JSON segments replace manual review of long-form video
The facts
- Pegasus 1.5 from TwelveLabs became available on Amazon Bedrock as of October 9, 2026
- Model analyzes videos up to two hours long in a single request, up from one hour in Pegasus 1.2
- New time-based metadata mode returns every matching segment as JSON with start/end times and requested fields
- Supports multimodal prompting with reference images, on-the-fly processing from S3 or base64 with no pre-indexing, and is served in US West (Oregon) via InvokeModel and InvokeModelWithResponseStream APIs
Amazon Bedrock now hosts Pegasus 1.5, a video language model from TwelveLabs. It turns long-form footage into structured, timestamped metadata. Earlier versions of Pegasus answered a single question about a clip. Pegasus 1.5 instead scans an entire video and returns every segment matching user-defined criteria as JSON. Each result carries start and end times, plus whatever fields the developer asked for: editorial segments, sports plays, speaker changes or brand appearances.
AWS highlights three new capabilities in the update. Multimodal prompting lets a segment definition pair a text description with a visual reference, to locate a specific product, logo or person. On-the-fly processing skips any ingestion pipeline or pre-indexing step, and sends video straight from Amazon S3. Extended processing capacity doubles the single-request video length, from one hour in Pegasus 1.2 to two hours.
Developers call the model as twelvelabs.pegasus-1-5-v1:0 in the US West (Oregon) region, on-demand. Access runs through the InvokeModel or InvokeModelWithResponseStream APIs. Input can be an S3 URI for long-form video, or a base64 string up to 25 MB for short clips.
Two analysis modes govern what comes back. General mode takes a text prompt and returns an answer, optionally constrained to a JSON Schema the developer supplies. Time-based metadata mode replaces the prompt with one or more segment definitions. Each specifies an id, a description of what a matching segment looks like, and the fields to extract, and returns a list of matching segments with their timestamps. AWS notes that data sent to the model stays within the AWS data boundary. An includeMetering flag adds token-count and video-duration fields for billing.
Why it matters
For teams building video search or compliance review, the shift from clip-by-clip Q&A to whole-video segment extraction removes the need for a separate indexing pipeline. The two-hour cap and S3-native input make it a plausible fit for broadcast and surveillance archives that previously needed custom toolchains to become queryable.
Primary source
AWS Machine Learning — aws.amazon.comWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.