This week, AWS rolled out several significant updates aimed at improving cost visibility, security, and governance for AI workloads. From granular cost allocation in Amazon Bedrock to the preview of Claude Mythos—a groundbreaking cybersecurity model—and the new Agent Registry for centralized agent management, these launches empower organizations to scale their AI projects with more control and insight. Below, we answer key questions about these announcements.
- How does the new cost allocation feature help teams manage AI spending?
- What is Claude Mythos and how is it available on Bedrock?
- What is the AWS Agent Registry and how does it improve governance?
- How can organizations access the Claude Mythos preview?
- What are the key capabilities of the Agent Registry?
- How does cost allocation work with IAM principals?
How does the new cost allocation feature help teams manage AI spending?
The latest Amazon Bedrock update introduces cost allocation by IAM user and role, allowing teams to tag IAM principals with attributes like team name or cost center. Once these tags are activated in the AWS Billing and Cost Management console, cost data flows automatically into AWS Cost Explorer and the detailed Cost & Usage Report. This gives finance and engineering leaders a clear view of who is using which foundation models and at what cost. For example, if multiple teams use Anthropic’s Claude or other models, you can now see exactly how much each team spends on inference. This eliminates guesswork and helps organizations optimize AI budgets as they move from experimentation to production. The feature also supports scenarios like tracking tool usage (e.g., Claude Code on Bedrock) and managing costs for multi-agent setups. Full setup details are available in the IAM principal cost allocation documentation.
What is Claude Mythos and how is it available on Bedrock?
Claude Mythos is Anthropic’s most advanced AI model to date, now available as a gated research preview on Amazon Bedrock through Project Glasswing. This model introduces a new class specifically designed for cybersecurity. It can identify sophisticated security vulnerabilities in software, analyze large codebases, and deliver state-of-the-art performance across cybersecurity, coding, and complex reasoning tasks. Security teams can use Claude Mythos to discover and address vulnerabilities in critical software before threats emerge. The preview is limited to allowlisted organizations, with Anthropic and AWS prioritizing internet-critical companies and open source maintainers. This launch marks a significant step in using AI proactively to harden software supply chains and reduce attack surfaces.
What is the AWS Agent Registry and how does it improve governance?
The AWS Agent Registry is a new preview service within Amazon Bedrock AgentCore that provides a private catalog for discovering and managing AI agents, tools, skills, MCP servers, and custom resources. Instead of duplicating existing agent capabilities, teams can now search the registry using semantic or keyword queries to find pre-built assets. The registry includes approval workflows for controlled sharing and tracks all changes via CloudTrail audit trails. This centralizes governance, reduces redundancy, and accelerates development by enabling reuse. The registry is accessible through the AgentCore Console, AWS CLI, SDK, and even as an MCP server that can be queried directly from IDEs. For organizations scaling AI across multiple teams, this is a powerful way to maintain consistency and oversight.
How can organizations access the Claude Mythos preview?
Access to Claude Mythos is currently limited to a gated research preview called Project Glasswing. Only allowlisted organizations can use the model on Amazon Bedrock. Anthropic and AWS are prioritizing companies that operate internet-critical infrastructure and maintainers of widely used open source projects. If your organization fits these criteria, you can request access through the Amazon Bedrock console or by contacting your AWS account team. The preview is intended to test the model’s capabilities in real-world cybersecurity scenarios before broader release. Given the model’s focus on vulnerability discovery and code analysis, early access may help security teams get ahead of emerging threats.
What are the key capabilities of the Agent Registry?
The Agent Registry offers several important features for managing AI agents at scale. First, it provides a private catalog where teams can publish and discover agents, tools, skills, MCP servers, and custom resources. Search is available via semantic and keyword queries, making it easy to find existing assets without reinventing the wheel. Second, the registry includes approval workflows, so administrators can control who can publish or use specific resources. Third, all actions are recorded in AWS CloudTrail, ensuring full auditability. The registry is accessible from the AgentCore Console, AWS CLI, SDK, and as an MCP server, allowing integration with development environments. This helps organizations reduce duplication, enforce governance policies, and maintain a clear inventory of their AI agent ecosystem.
How does cost allocation work with IAM principals?
Cost allocation by IAM user and role in Amazon Bedrock leverages existing AWS tagging mechanisms. You first tag IAM users or roles with custom attributes such as team, cost center, or project. Then, in the Billing and Cost Management console, you activate those tags for cost allocation. Once activated, all Bedrock inference costs incurred by those IAM principals are tagged accordingly. The tagged cost data then appears in AWS Cost Explorer and the detailed Cost and Usage Report, enabling you to filter and group expenses by team or cost center. This works regardless of whether you are using single models, multi-agent systems, or tools like Claude Code. Setting up this feature is straightforward; full instructions are in the IAM principal cost allocation documentation.