Claude Managed Agents Pricing: A Complete Breakdown for 2026

Claude Managed Agents pricing looks simple on paper: tokens plus a runtime fee. In practice, most teams underestimate the bill because the runtime meter, the idle-time exemption, and the list of discounts that quietly stop applying once you move from the Messages API to a managed session aren't obvious until you've read the fine print. This is for engineering and platform teams sizing out what a Claude Managed Agents deployment will actually cost before committing budget to it.

What Claude Managed Agents actually bills you for?

Anthropic bills Claude Managed Agents sessions on exactly two dimensions, according to its official pricing documentation: the tokens the agent consumes, and the time the session spends in a running state.

Every input token, output token, cache write, and cache read is billed at the same per-model rate as the standard Claude API. There's no markup for running inside a managed session. On top of that, session runtime is metered separately at $0.08 per session-hour, tracked to the millisecond, and it only accrues while the session status is running. Time spent idle (waiting on your next message or a tool confirmation) doesn't count, and neither does time spent rescheduling or after the session is terminated.

Claude Managed Agents pricing plans

There's no separate subscription tier for Managed Agents itself. You pick a model, and the session runtime fee is the same $0.08/hour no matter which model is running underneath it. Volume discounts and custom rate limits are negotiated directly with Anthropic's sales team for high-volume deployments.

What's included versus what costs extra?

A few things catch teams off guard once they start running real sessions:

• Idle time is free. A session waiting on a human response, a tool approval, or a scheduled resume doesn't burn runtime charges, only active execution time does.
• Batch API discounts don't apply. The 50% batch discount that applies to standard Messages API calls has no equivalent for Managed Agents sessions, since sessions are stateful and interactive rather than queued.
• Fast mode and data residency multipliers don't apply either. Fast mode's premium pricing and the 1.1x inference_geo: "us" multiplier are both Messages API request fields, so neither one carries over to session billing.
• Web search still costs extra. If a session triggers a web search, it's billed at the standard $10 per 1,000 searches on top of the tokens the search results generate.
• Session runtime replaces container-hour billing. If your session uses code execution, you're not billed separately for container hours the way you would be on the standalone Messages API. That cost is folded into the session runtime fee instead.

Worked example

Anthropic's own documentation walks through a one-hour coding session on Claude Opus 5 that consumes 50,000 input tokens and 15,000 output tokens:

For a higher-volume comparison: processing 10,000 support tickets on Claude Haiku 4.5, at roughly 3,700 tokens per conversation, costs approximately $37 total. The runtime fee barely moves the needle on short, high-throughput sessions; it starts to matter more on long-running agents that sit in a running state for extended periods, like an agent monitoring a queue or working through a multi-hour research task.

Claude Managed Agents pricing vs TrueFoundry Agent Harness

Anthropic's runtime fee is straightforward for a single Claude-only agent. It gets less straightforward once you're running many agents across a team, need models beyond Claude, or need the whole thing deployed somewhere other than Anthropic's managed cloud.

TrueFoundry is open-sourcing Agent Harness, so the harness itself carries no license fee: you self-host it on your own infrastructure, in your own cloud or on-prem. It is a managed agent harness built on top of TrueFoundry's AI Gateway and MCP Gateway. The only harness that sits inside a full 4-layer governance platform: AI Gateway → MCP Gateway → Agent Gateway → Agent Harness. Includes both a no-code builder UI and a pro-code API.

In TrueFoundry Agent Harness - API keys, OAuth tokens, and provider credentials never appear in agent definitions. They live in the gateway layers. Agent builders reference model names and tool names only. Platform teams configure access once; rotation is centralized. Key features include:

• End-to-end traces per agent run: every LLM call, tool call, sandbox execution, subagent call.
• Cost, tokens, and latency per step not just per run.
• Inherits AI Gateway analytics: org-wide model usage, request logs, per-user/team cost breakdowns.
• MCP Gateway traffic included.
• OpenTelemetry export: plug into any APM.
• Prometheus/Grafana: ops-grade alerting, dashboards.
• One pane of glass across model traffic, MCP traffic, and agent traffic.

Why this matters: When an agent run costs $4 instead of $0.40, you need to know which tool call caused the blowup, which subagent triggered the cascade, and which user is responsible. Only TrueFoundry gives you that drill-down by default.

Key advantage: TrueFoundry's model-level budgeting means organizations can actually control model spend, which Anthropic and LangChain don't offer natively.

Total cost of ownership considerations

The per-session math above is only part of the real bill. Here's how the two stack up once you look past a single session, and where that comparison tends to favor which platform:

• Multi-agent scale changes the math. Claude Managed Agents charges $0.08 per session-hour on top of tokens, for every agent, every session. That's trivial for one agent, but multiply it across dozens or hundreds of concurrent sessions running continuously and it becomes a real, recurring infrastructure line item, separate from token spend, that scales with agent count rather than with actual compute used. Agent Harness carries no per-session runtime fee. Since it's open source, you pay for the infrastructure you host it on, which scales with your own compute costs rather than a per-agent vendor meter.
• Vendor lock-in has a cost even when it isn't billed directly. Claude Managed Agents only runs Claude models, so evaluating another provider later means a rebuild, not a config change. Agent Harness routes through the AI Gateway to 1,000+ LLMs across providers, so switching or mixing models is a config change, not a re-architecture, and model-level RBAC and budgets come with that access rather than being a separate build.
• Credential handling gets more expensive to manage as you grow, not just to run. Claude Managed Agents registers credentials in a vault per user per session, so the operational overhead multiplies with headcount and agent count. Agent Harness centralizes MCP credentials in the MCP Gateway with OAuth and automatic rotation: admins update a credential once and it applies to every agent using it, which costs less to operate at scale even though you're the one hosting it.
• Data residency has real requirements behind it, not just preference. Claude Managed Agents' managed cloud is US-only in beta, with EU availability planned post general availability, and no on-prem option. If air-gapped or fully on-prem deployment is a hard requirement today, that's not something Anthropic's managed offering currently supports. Agent Harness already deploys as SaaS, self-hosted, or on-prem in your own region, so that requirement is solved now rather than dependent on Anthropic's roadmap.

Line for line, the costs that compound quietly on Claude Managed Agents at multi-agent scale, the runtime meter, the single-model ceiling, per-session credentials, and the residency gap, are the ones Agent Harness is built to remove.

Conclusion

Claude Managed Agents' pricing is genuinely simple at the level of a single session: tokens plus $0.08 an hour, with idle time free. Where it gets more expensive is at scale, across many agents, multi-model support, centralized credential management, enterprise governance, observability, deployment flexibility, and infrastructure ownership become increasingly important, especially when operating dozens or hundreds of agents across teams.

For these scenarios, TrueFoundry Agent Harness is the stronger long-term choice. Rather than locking teams into a single model provider, it provides an open-source, vendor-agnostic platform for governing and deploying AI agents built with any framework. Combined with the AI Gateway and MCP Gateway, TrueFoundry enables organizations to manage model access, credentials, security policies, observability, and deployments from a single control plane, whether running in the cloud, on-premises, or inside a private VPC.

FAQ

What is Claude Managed Agents pricing?

‍Claude Managed Agents bills on two dimensions: standard Claude API token rates for everything the agent reads and generates, plus a flat $0.08 per session-hour for runtime, metered only while the session is actively running. There's no separate subscription fee for the managed runtime itself.

Is Claude Managed Agents free?

‍No. It uses the same token pricing as the standard Claude API, plus the session runtime fee. New accounts get a small amount of free credit to test it, but there's no ongoing free tier for production use.

Does Agent Harness cost anything to run?

‍The harness itself is open source, so there's no license fee. You pay for the infrastructure you self-host it on and for whatever model or MCP tool calls your agents make, the same way you would with any self-hosted framework.

Does TrueFoundry support MCP and AI agents?

‍Yes. TrueFoundry includes an MCP Gateway, an Agent Gateway, and an MCP and Agents Registry with tool-level access control, so agents from any framework can be deployed and governed through one control plane.

Does TrueFoundry integrate with my existing observability stack?

‍Yes. The platform is OpenTelemetry-compliant and plugs into Grafana, Datadog, Prometheus, or your existing stack, tracing every request from prompt through tool and model execution.

Related reading

• What Is an Agent Harness?: the infrastructure layer underneath managed and self-hosted agents alike
• Claude Managed Agents vs Hermes Agent: how a fully managed runtime compares to a self-hosted alternative
• LangGraph Pricing: A Complete Breakdown for 2026: how a different agent framework's hosting costs stack up
• AI Agent Observability: Monitoring and Debugging: what to track once you're running agents in production