Amazon Bedrock Projects API: finally assigning GenAI costs by team and project

 |  Eric Pinet

Introduction

AWS has introduced the Projects API in Amazon Bedrock (Mantle engine, compatible with OpenAI), allowing organizations to isolate GenAI workloads by project, team, or environment. Each project comes with its own IAM permissions and cost allocation tags, providing granular visibility into inference spending.

This marks AWS’s first concrete response to a growing challenge for FinOps teams: controlling and allocating the costs of generative AI, now the fastest-growing category of cloud spending in 2026.

What’s changing?

Until now, all Bedrock inference requests within a single AWS account were grouped into one cost bucket in Cost Explorer. As a result, it was nearly impossible to attribute spending by team, product, or environment without building complex custom instrumentation.

The Projects API introduces a “project” concept within Bedrock, acting as a logical container for GenAI workloads. In practice, a project can receive:

  • Custom AWS tags for cost allocation (for example: Team=data, Project=client-chatbot)
  • Dedicated IAM policies to control access by application or team
  • Complete isolation from other projects within the same AWS account

Technically, projects run on Bedrock’s Mantle engine, which exposes APIs compatible with OpenAI (Responses API and Chat Completions API). For organizations already integrated with OpenAI, migration is therefore almost seamless.

Why this matters for FinOps

Generative AI spending has become the fastest-growing cost category on AWS in 2026. Yet most organizations still lack visibility into who is consuming what within Bedrock.

The Projects API changes that dynamic in four key ways:

  1. Chargeback and showback for AI costs: allocate inference spending to the correct cost centers without building internal tracking tools.
  2. Team-level budgets and alerts: configure AWS Budgets for individual Bedrock projects and trigger alerts before spending exceeds limits.
  3. Waste detection: identify teams using premium models (such as Claude Opus) for tasks that do not require them.
  4. Model decision management: compare the cost per task across models and justify moving to more economical alternatives.

Supported models as of March 6, 2026 in us-east-1

For more information: