Summary

  • The September 10 integration packages Marengo Embed 3.0 with managed ingestion, indexing and retrieval; it is not TwelveLabs' first appearance on Bedrock.
  • Storage, search calls and model use have different billing units. This native multimodal path accepts text queries and retrieves results rather than generating an answer.

A cheap model is not necessarily a cheap search service. Before an enterprise can judge whether video embeddings find useful moments in its footage, it needs somewhere to index them and a way to turn a query into ranked results. TwelveLabs and AWS are moving that assembly work into a managed product. The commercial question shifts from what a model costs to what a usable search workload costs.

AWS announced Marengo Embed 3.0's general availability in Amazon Bedrock Knowledge Bases on September 10. TwelveLabs describes the resulting offer as a managed endpoint: AWS provides infrastructure, connectors and orchestration, while Marengo supplies embeddings and TwelveLabs contributes the retrieval logic. Customers no longer need to provision the vector database and assemble that search stack themselves.

This is a change in packaging, not the beginning of the relationship. AWS made Marengo 2.7 and Pegasus 1.2 available on Bedrock in July 2025. Those earlier models and their regional arrangements should not be confused with the new knowledge-base integration, or with separate Marengo 3.5 announcements.

Read the meter, not just the model

AWS's public managed knowledge-base price table lists index storage at $5 per GB of raw data per month and standard retrieval at $1 per 1,000 API calls. The built-in managed parser, embedding model and reranker carry no extra charge in that table. A footnote matters: choosing one's own embedding or reranking model adds provider charges. Optional Gateway invocation and enabled observability also have their standard charges.

The same page illustrates direct Marengo embedding with ten videos totaling 100 minutes: 6,000 seconds at $0.00070 costs $4.20. That is a model-use example, not the monthly price of a searchable archive. Duration does not tell the buyer how many raw gigabytes it holds or how often users will query it.

Nor does a compact representation settle the storage bill. AWS describes Marengo 3.0's embeddings as 512-dimensional vectors; its managed index price is denominated in raw data. Smaller vectors cannot, by themselves, establish a saving against that denominator. No actual invoice or total-cost comparison was examined for this report.

Search has a narrower contract than generation

The native multimodal guide sends media directly to the embedding model rather than first reducing it to text chunks. Audio and video use configurable segmentation. That can retain signals a text-only representation would omit, but it is not a claim of lossless understanding or perfect retrieval.

The interface also has clear limits. This knowledge-base path supports Retrieve, not RetrieveAndGenerate, and accepts text queries rather than image queries. A matching clip is a search result, not an automatically composed answer or proof that an event happened as a user interprets it. Broader capabilities and image-input pricing for the underlying model do not expand this particular interface.

Some operating choices remain with the customer. AWS requires a separate S3 destination for multimodal processing, alongside the source bucket. It attempts to remove transient processing data and recommends narrowly scoped expiry; applying deletion to the entire bucket or broad service prefix would remove useful content. Regional invocation paths and the permissions for ingestion and queries also need checking. TwelveLabs' claim that data stays in the customer's account should not be read as an independently verified promise of single-region processing.

The opportunity is a more practical evaluation: useful retrieval can be considered without first commissioning a substantial search build. The suppliers' claims about almost immediate results and months avoided remain marketing claims. Their demonstration is not an enterprise benchmark.

Sources: TwelveLabs' integration announcement; AWS's September 10 release and walkthrough; native multimodal constraints; AWS pricing; the July 2025 model launch.