Anthropic Just Launched Managed Agents. Let's Talk About How We're Going to Pay for This

Anthropic's Claude Managed Agents introduces a three-dimensional billing model — tokens, session runtime, and per-tool charges — that exposes a fundamental gap in existing FinOps practices. The author breaks down the pricing structure, highlights the attribution problem for autonomous workloads, and draws parallels to familiar cloud cost patterns like Reserved Instances. Key lessons include: AI agent costs are multi-dimensional and don't map to existing dashboards, autonomous workloads have no natural cost ceiling, model selection should be evaluated on cost-per-outcome not cost-per-token, and prompt caching offers up to 90% savings on repeated context. The post argues that current FinOps frameworks built for VMs and serverless aren't equipped for agentic workloads, and new primitives are needed.