Amazon Bedrock’s Pegasus 1.5 turns two-hour videos into timestamped JSON

Amazon Web Services has made TwelveLabs’ Pegasus 1.5 video language model available on Amazon Bedrock. According to the AWS blog post, it scans a video of up to two hours and returns every segment matching your criteria as timestamped JSON.

AWS says the availability was first announced on October 9, 2026. The detailed technical walkthrough followed on October 19.

Describe the moments, get segments back

Earlier Pegasus versions answered a question about a clip. Pegasus 1.5 adds a second mode, time_based_metadata. The developer sends one or more segment definitions, each with an ID, a plain-language description and the fields to extract (string, boolean, number, integer, array or timestamp). The model returns each matching stretch of video with start and end times and those fields filled in. A single request can carry several definitions, for example editorial sections plus on-screen brand appearances, and each result is tagged with the definition it matched.

The default mode, general, still takes a text prompt and can be forced to follow a JSON Schema. Reference images are new: a definition can pair text with a picture of a product, logo or person, up to 30 images per request, so the model can locate that item and report what happens around it.

The practical limits, per AWS:

• Videos up to 2 hours, up from one hour in Pegasus 1.2.
• Up to 2 GB from an Amazon S3 location, or 25 MB as a base64 string.
• Resolution from 360×360 to 5184×2160, and aspect ratios between 1:1 and 2.4:1.
• No pre-indexing step, and the data stays inside the AWS data boundary.
• Full support for English, partial for 12 other languages including Arabic, Chinese, Japanese and Spanish.

Calls go through InvokeModel or InvokeModelWithResponseStream using the model ID twelvelabs.pegasus-1-5-v1:0. The sample code uses the global cross-Region profile, global.twelvelabs.pegasus-1-5-v1:0, and AWS suggests a longer read timeout, 900 seconds in its example, for segmentation jobs.

Oregon only, billed outside Bedrock

The blog post lists a single Region, US West (Oregon). A separate AWS announcement says access also runs through US and global cross-Region inference profiles but gives no Region list. The post does not state a price. It points to the Bedrock pricing page and says charges come through AWS Marketplace, so they show up on the AWS bill and in Cost Explorer under TwelveLabs, not under Bedrock. Teams that budget per service should expect that to look different from other Bedrock models.

Access needs an AWS account, model access enabled in the Bedrock console and IAM permissions for the two invoke actions. AWS also describes tracking cost per application through application inference profiles.

How it compares with general-purpose video models

The closest alternative for many teams is a general multimodal model such as Google’s Gemini, which can also reason over video. Google’s developer documentation says Gemini samples video at 1 frame per second and uses roughly 100 tokens per second of video at default resolution, about 300 at high resolution. Models with a 1M-token context can take videos up to 3 hours at default resolution, but only 1 hour at high resolution. Timestamps are referenced in MM:SS form in the prompt.

Those are different designs. Gemini’s video handling is bounded by the context window, and the user writes the prompt that asks for timestamps. Pegasus 1.5 makes the segment list the product: a schema in, a typed list of segments out. AWS’s own sample shows the cost of that depth, with a segmentation run of roughly nine and a half minutes of sports footage consuming 249,408 input tokens. Whether that is cheaper or more accurate than a general model is not something either vendor’s material establishes.

Pegasus is also not new to Bedrock. According to AWS, TwelveLabs models became fully managed on Bedrock in July 2025, with Pegasus 1.2 available in Oregon and Ireland through cross-Region inference. This release raises the length ceiling and changes the output, but the Region footprint listed in the new post is narrower than what Pegasus 1.2 launched with.

What AWS has not shown

The post contains no accuracy figures, benchmarks or latency numbers for segmentation. The sports and canoe outputs are illustrative samples, not evaluations, and the claims about speeding up highlights or cataloging archives are AWS’s and TwelveLabs’ framing. AWS itself says people stay in the loop, reviewing and refining the generated metadata.

Open questions for buyers are therefore concrete: how well segment boundaries hold up on a full two-hour file, how the Marketplace pricing scales with video length (the optional includeMetering flag reports analyzed seconds and output tokens), and when Regions beyond Oregon arrive. The source gives no timeline for any of them.