TwelveLabs Pegasus 1.5 on Bedrock Adds Two-Hour Video Segmentation

Bedrock availability began on October 9, not October 19

Amazon Web Services published its detailed Pegasus 1.5 tutorial on October 19, 2026. The post itself points back to an October 9 Bedrock availability announcement, while the live model card gives October 8 as the model launch date. The current event is a substantial implementation disclosure for an already available model, not a second launch.

Pegasus 1.5 is TwelveLabs’ video-to-text model for summaries, detailed descriptions and structured responses about video. On Bedrock it is active and offered on demand through bedrock-runtime. AWS identifies the base model as twelvelabs.pegasus-1-5-v1:0, but production calls must use a cross-Region inference profile rather than that base ID.

The main change is whole-video segmentation into timestamped JSON

In the default general mode, a builder sends a text prompt and receives text, or JSON when a response schema is supplied. In time_based_metadata mode, the builder defines the segment types that matter—such as editorial sections, speaker changes, sports plays or brand appearances—and requests named fields. The response is a segments list with start and end times, requested fields and a definition identifier.

AWS says Pegasus 1.5 can process videos up to two hours in one request, double the one-hour limit it attributes to Pegasus 1.2. That is a capacity statement, not evidence that every relevant event will be found or bounded accurately. Evaluate missed segments, false matches, boundary drift, field accuracy and repeatability on annotated videos from the intended domain.

Reference images and direct video input remove an indexing step

A segment definition can combine text with a visual reference, allowing a request to ask where a particular logo, product, person or scene appears. AWS documents up to 30 reference images in one request. General mode places references beside the prompt; segmentation mode attaches them to the relevant definition.

The video itself can be read from an S3 URI or carried as a base64 string. Inline video is limited to 25 MB, so long-form work will normally use S3. AWS presents this as on-the-fly processing with no required pre-indexing or ingestion pipeline. That can simplify a first integration, but repeated analysis of the same library may still make caching, durable metadata and a separate search index economically useful.

Cross-Region inference is mandatory

The live model card says Pegasus 1.5 does not support in-Region on-demand inference. US geographic requests use us.twelvelabs.pegasus-1-5-v1:0 and can route among us-east-1, us-east-2 and us-west-2. Global requests use global.twelvelabs.pegasus-1-5-v1:0 and can route across supported AWS Regions.

A Global profile is therefore not an appropriate default when residency or transfer boundaries require a named geography. Even with a cross-Region profile, AWS says an S3 bucket must match the source Region where the request is sent; a bucket in another Region produces a validation error. Record both the source endpoint and inference profile in architecture and compliance reviews.

The API has large limits and explicit completion signals

AWS lists a 261,120-token context window and 98,304 maximum output tokens. InvokeModel and InvokeModelWithResponseStream are the documented invocation paths. Responses include input and output token counts and a stopReason; a length stop means generation ended at maxOutputTokens and must not be treated as a complete analysis.

Those limits do not establish reliable two-hour comprehension. Long inputs can contain dense cuts, overlapping speakers, small text and repeated visual motifs. Cap output intentionally, check stopReason, reject malformed or incomplete segment sets, and compare returned intervals against human-labeled fixtures rather than assuming a large context guarantees temporal recall.

Pricing has two metered dimensions but needs a live quote

AWS’s TwelveLabs pricing table exposes a price per second of video input and a price per one million output tokens for Pegasus 1.5. The page says those Pegasus rates apply to both US and Global cross-Region inference. The model card adds that this is an AWS Marketplace product whose charges appear under the model provider in Cost Explorer rather than under Amazon Bedrock.

The numeric price cells are injected client-side and did not resolve in the public response reviewed by AccessAllGPT, so we do not quote a rate. Before a trial, capture the live console or pricing-API values, currency and timestamp. Estimate with actual video duration and expected output, then include S3 storage and requests, transfer, retries, reprocessing, human review and any downstream indexing.

What video-AI buyers should do next

Trial Pegasus 1.5 on a stratified, rights-cleared sample that represents video lengths, languages, cuts, audio quality and event density. Build ground truth for both existence and start/end boundaries. Compare general mode and time-based metadata on missed events, false positives, timestamp error, field validity, truncation, latency and total cost per accepted video.

Adopt when the tested profile satisfies residency, quality and complete-video economics. Constrain it to assisted review when misses or boundary errors carry moderate consequences. Wait if numeric pricing, quota or required geography cannot be confirmed. Reject unattended use for consequential identification or enforcement when a plausible segment error can harm a person or create an irreversible action.

Disclosures

AccessAllGPT received no AWS or TwelveLabs account, credentials, credits, early access, briefing, video, benchmark data, review, payment or compensation for this article. Neither company sponsored, reviewed or endorsed it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Amazon Web Services or TwelveLabs. Publication-wide relationships are listed on the disclosures page.