In-depth analysis · About 6 min read
AWS details Pegasus 1.5 video analysis on Bedrock
The October 19 guide explains October 9 availability, timestamped JSON and video analysis up to two hours.
- Source published
- Published here
- Source checked
AWS published guidance for TwelveLabs Pegasus 1.5 on Bedrock. Its October 19 post references the October 9 availability announcement: this is a follow-up, not a launch today.
Listen to the news
Read this article aloud or keep listening to the news.
Preparing the reader…
Device voice
Speech is handled by your browser or device. AISubHub does not send the text to an audio API. Voice availability and playback with a locked screen vary by device.
Resuming may repeat a short phrase.
At a glance
- The model can find user-defined video segments and return timestamps and structured fields; reference images can help describe a target.
- AWS lists videos up to two hours, input from S3 or small inline clips, and availability in the US West Oregon region.
A follow-up guide, not another launch
AWS’s October 19 article explains an availability announcement made on October 9. Its substantive subject is how Pegasus 1.5 turns a whole video into segments defined by the application, rather than merely answering one question about a clip. AWS describes no pre-indexing requirement: a request can refer directly to an S3 object or carry a short video inline. The claim that processing stays within the AWS data boundary belongs to the vendor, rather than to an audit performed by AISubHub.
The guide lists on-demand inference in US West (Oregon), us-west-2, with model ID twelvelabs.pegasus-1-5-v1:0. It uses bedrock-runtime with InvokeModel or InvokeModelWithResponseStream. Its sample also uses a global inference-profile prefix and mentions a US alternative. That example should not be read as proof of universal regional availability; deployment teams still need the current model card and permitted processing geography.
Two analysis modes require different payloads
General mode is the default. An inputPrompt asks a question, and the answer arrives in message. When responseFormat.type is json_schema, the schema goes under jsonSchema and message contains a JSON string for downstream parsing. In contrast, time_based_metadata mode sets responseFormat.type to segment_definitions. The definitions replace inputPrompt, which the guide says to leave out in this mode.
Each segment definition has an id, a description of a matching event, and requested fields with names, types and descriptions. A broadcaster might request editorial sections with a title, summary and topic list; another definition can find brand appearances. Multiple definitions share a segments result list, with _definition_id identifying which criterion matched. Analysis: these definitions are a reusable annotation policy, so editorial teams should agree what counts as a segment before expecting comparable results across a library.
Timestamps and visual references need explicit handling
Segments carry start_sec, end_sec and the requested fields. Despite their names, those time values are not always numeric: segmentTimeFormat can request seconds, hh:mm:ss or hh:mm:ss.fff. The response also includes definition_errors. Analysis: code should check that list, normalize the chosen timestamp representation and preserve definition identity before loading results into a search index. A partly processed request should not silently become a complete catalogue.
Reference images can narrow a description to a particular product, logo or person. General mode places them in promptV2.mediaSources with the question in promptV2.inputText; segmentation attaches mediaSources to each definition. Each image needs a name, media_type:image and exactly one source, base64String or an S3 location. A request can contain up to 30 reference images. The guide demonstrates a canoe reference; it is an illustration, not an independent product-recognition benchmark.
Input limits are part of the design
The guide allows videos up to two hours, compared with one hour in Pegasus 1.2. S3 inputs can be up to 2 GB, while inline base64 inputs are limited to 25 MB. It specifies resolution from 360×360 to 5184×2160, aspect ratios from 1:1 to 2.4:1 or their inverses, and FFmpeg-supported formats. These conditions determine whether existing archive assets can be sent directly or require preprocessing.
An S3 request includes uri and bucketOwner. Prerequisites include AWS and model access, Python 3.10 or later with Boto3, and permissions for the invoke actions. The article describes normal signed credentials and an optional Bedrock bearer-token route requiring bedrock:CallWithBearerToken; it recommends short-lived keys for production. These are cloud integration requirements, not access included in a personal video subscription.
Watch completion, usage and billing separately
The sample uses a 900-second read timeout, a 10-second connection timeout and no SDK retries, showing that long video jobs require deliberate client settings. Those values are example configuration, not an observed latency promise. Responses expose stopReason; length means maxOutputTokens ended generation early. Streaming returns message chunks, but seeing chunks is not equivalent to receiving a complete set of segment metadata.
With includeMetering:true, the non-streaming response adds outputTokenCount and inputVideoSecond for usage tracking. CloudWatch provides invocation, output-token, latency and client-error metrics; tagged application inference profiles can separate application usage. AWS says Marketplace charges appear under TwelveLabs on the AWS bill and in Cost Explorer, rather than under Amazon Bedrock. This guide does not establish a total price for a specific video collection, and AISubHub has not measured one.
Analysis: evaluate metadata, not just fluent descriptions
The article gives news archives, sports highlights, contextual recommendations and brand-safety review as applications. A useful pilot would freeze definitions, annotate representative clips by hand, then compare missed events, false matches and boundary timing. Reference-image matching and text-only matching should be assessed separately. A polished summary or valid JSON is insufficient evidence that all relevant moments were found. Include negative clips with no target, visually similar distractors and ambiguous transitions; classify errors by definition, image matching and timing rather than judging only fluent summaries.
English is fully supported, while the listed Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish, Thai and Vietnamese support is partial. A translated interface does not imply equal model performance in every language. AWS itself keeps people in the review loop. Our assessment is therefore a workflow proposal, not a claim of tested accuracy, latency, copyright clearance or savings: generated timestamps and labels should remain reviewable before production use.
Why it matters
For archive editors or sports teams, segment-level metadata could make existing footage easier to review and organize. A useful pilot would compare generated timestamps with human annotations, checking missed scenes and false matches before metadata enters a production search system.
Limits and availability
The article describes cloud API access requiring an AWS account, permissions and model access, not a consumer video subscription. AWS says billing runs through Marketplace under TwelveLabs. Language support varies, and human review remains necessary; we have not independently tested accuracy, latency or total processing cost.
Sources and evidence
Source published: Oct 19, 2026 · Source checked: Oct 20, 2026
AI-assisted editorial articles checked against linked primary sources. Vendor claims are attributed; inclusion is not endorsement.