Pegasus 1.5 from TwelveLabs is now available in Amazon Bedrock

TwelveLabs' Pegasus 1.5, a video language model that converts long-form video into structured, timestamped metadata, became available in Amazon Bedrock on October 9, 2026, according to AWS' Machine Learning Blog. The model runs on the bedrock-runtime endpoint under the ID twelvelabs.pegasus-1-5-v1:0 in US West (Oregon), us-west-2, and is called through the on-demand InvokeModel or InvokeModelWithResponseStream APIs. Requests can point at video stored in Amazon S3 or embed a short clip inline, and no pre-indexing or ingestion pipeline is required.
The central change from Pegasus 1.2 is the reversal of how a request is framed. Previously the pattern was to give the model a clip and a question and receive an answer. Pegasus 1.5 instead accepts a set of segment definitions describing moments that matter to a workflow, such as editorial sections, sports plays, speaker changes or brand appearances, and returns every matching part of the video as JSON with start and end times and the requested fields filled in.
Each request operates in one of two modes selected by the analysisMode field. General mode, the default, takes a text prompt in inputPrompt and returns a text answer in message, or JSON when a schema is supplied. Time-based metadata mode takes segment definitions in responseFormat, replaces inputPrompt entirely, and returns a segments list. Each definition carries an id, a description of what a matching segment looks like, and fields to extract, where each field has a name, a description and a type drawn from string, boolean, number, integer, array or timestamp.
Segment timestamps are formatted according to segmentTimeFormat: seconds returns numbers suitable for arithmetic, while hh:mm:ss and hh:mm:ss.fff return readable strings. Responses also include a definition_errors list that stays empty when every definition processed. Multiple definitions can be sent in one request, for example one for editorial sections and another for on-screen brand appearances, and each returned segment is tagged with _definition_id so results can be grouped. Because the schema is defined once, the same definitions can be applied across an entire video library for consistent metadata.
Video can be supplied two ways. An s3Location takes the object's uri plus the bucketOwner account ID and is suited to long-form content; a base64String embeds the video in the request body up to 25 MB and suits short clips. Pegasus 1.5 handles videos up to two hours long in a single request, up from one hour in Pegasus 1.2. Alongside segmentation, it retains video-to-text generation, analysis of objects, on-screen text and numerical content, temporal grounding, combined visual, audio and textual understanding, and multilingual support, with full English support and partial support for Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish, Thai and Vietnamese.
Multimodal prompting lets a request include reference images, so a user can show a product, logo or person rather than describe it. In general mode, images sit in promptV2.mediaSources next to the question in promptV2.inputText. In segmentation mode they attach to the mediaSources of a segment definition, letting that definition match on a visual reference as well as a text description. Each image needs a name, a media_type of image, and exactly one source, either inline base64 or an S3 location. A single request can carry up to 30 reference images. Obvious use cases include visual brand monitoring, product placement analysis and image-guided content search.
Tuning parameters include temperature from 0 to 1 with a default of 0.2, and maxOutputTokens from 512 to 98,304. In general mode, responseFormat can be set to json_schema with a JSON Schema under jsonSchema, and the answer arrives in message as a JSON string that downstream code can parse. Responses report input_tokens, output_tokens, and a stopReason; a value of length signals the answer hit maxOutputTokens before finishing. Setting includeMetering to true adds two billing-usage fields to non-streaming responses: outputTokenCount and inputVideoSecond, the analyzed video's duration in seconds.
Callers need an active AWS account with Bedrock access, Python 3.10 or later with Boto3, IAM permissions for bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream, and model access enabled in the Bedrock console. Authentication can use standard AWS credentials, or a Bedrock API key set in the AWS_BEARER_TOKEN_BEDROCK environment variable, which additionally requires bedrock:CallWithBearerToken; AWS recommends short-term keys generated with the aws-bedrock-token-generator package for production. Requests can target global Cross Region Inference or the US geographic endpoint by swapping the endpoint prefix. AWS advises a longer read timeout, around 900 seconds in its sample, for lengthier invocations such as video segmentation. For cost tracking by application, an application inference profile can be created from the model, tagged, and called by its ARN instead of the model ID.
My reading is that the shift from question-answering to schema-driven extraction is the significant part for developers. A fixed schema plus a segments list resembles an indexing job that can feed search, editorial tooling or ad verification directly, rather than a chat response needing post-processing. The two-hour ceiling and per-request S3 access simplify pipelines but also mean each call does substantial work, which is why output tokens run to 98,304 and token budgets are reported back. The end of the source shows a segmentation response with input_tokens of 249,408 and output_tokens of 802 for a sports clip, illustrating how input-heavy these calls tend to be relative to output.
Why it matters: Teams building video search, media monitoring or content operations get a managed way to turn whole video libraries into timestamped JSON without running their own indexing infrastructure, with processing and data held inside the AWS boundary. Product teams should expect to design segment schemas up front and budget around video duration and input-token volume, while developers evaluating vision-language alternatives should note the reference-image matching and the two-hour single-request limit as the practical differentiators.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
TwelveLabs Pegasus 1.5 is now available in Amazon Bedrock. The video language model turns long-form video into structured, timestamped metadata: define the moments that matter and get every matching segment back as JSON, with new video segmentation, multimodal prompting with reference images, and on-the-fly processing of up to two hours of video.