Claude Managed Agents: Pricing and How It Works

Q: Which models does it support?

You set the model on the agent, for example claude-opus-5-5, and self hosted sandboxes support every Claude model available in Managed Agents. Fast mode is available on supported Opus models at premium token rates.

Claude Managed Agents is Anthropic's hosted agent service: you describe an agent (model, system prompt, tools, MCP servers and skills), and Anthropic runs the agent loop, the sandbox where its commands execute, and the session history for you. You pay normal token rates plus $0.08 for every hour a session is actively running. It is still in beta, it is turned on by default for every Claude API account, and it is the closest thing Anthropic sells to "Claude Code as a backend service."

This guide explains what the service actually hosts, how the bill is calculated, when the self hosted sandbox option makes sense, the limits you will hit, and the open source projects people compare it with. Everything here is checked against Anthropic's documentation as of October 23, 2026.

What's new (October 23, 2026): Anthropic launched Claude Models on Google Cloud (Anthropic managed) on October 22. It is the Claude API run by Anthropic and reached through your Google Cloud project, with Google Cloud IAM for access and Google Cloud Marketplace for billing, and Managed Agents (agents, environments, sessions, vaults, memory stores, scheduled deployments and self hosted sandboxes) works on it. Source: Claude Platform release notes.

What Claude Managed Agents is

Anthropic now offers two ways to build with Claude. The Messages API gives you the model and nothing else, so you write the loop that calls tools, feeds results back and decides when to stop (our guide on building an AI agent from scratch walks through exactly that). Claude Managed Agents gives you a finished agent harness instead. Anthropic's overview page describes it as a configurable harness running in managed infrastructure, aimed at long running and asynchronous work.

The service is built on four objects:

Agent: a reusable, versioned configuration holding the model, system prompt, tools, MCP servers and skills.
Environment: where tool calls run. Either an Anthropic cloud sandbox or a sandbox you host.
Session: one running instance of an agent inside an environment, working on one task.
Events: the messages that flow between your app and the agent, streamed back over server sent events.

When you send a message, the platform provisions a sandbox, runs the agent loop, executes tool calls inside the sandbox, streams events to you, and finally emits a session.status_idle event when the agent has nothing left to do. Conversation history, files and sandbox state are kept on Anthropic's side, so a session can pause and pick up again later.

Built in tools

Adding the agent_toolset_20260401 tool type switches on eight tools at once: bash, read, write, edit, glob, grep, web fetch and web search, according to the tools reference. You can disable any of them individually, restrict web search and web fetch to a list of allowed or blocked domains, and connect external tools through the Model Context Protocol. Tool output over 100,000 characters is written to a file in the sandbox and the model gets a preview plus the path.

Features beyond the loop

The extras are what separate this from a plain sandbox. Memory stores are folders of text documents mounted into the sandbox so an agent can carry preferences and project notes across sessions. Vaults inject credentials at egress so secrets never sit in the prompt. Scheduled deployments run an agent on a cron expression with a time zone, which replaces your own scheduler. Multiagent orchestration lets one agent hand work to subagents, write a dynamic workflow that fans out across many agents, or consult a stronger advisor model mid turn. Skills load the same way as in Claude Code (see our Agent Skills guide), including from a mounted GitHub repository's .claude/skills folder.

How to start a session with the API

The quickstart needs a Claude Console account, an API key and the SDK (pip install anthropic; we confirmed the beta.agents, beta.environments and beta.sessions resources exist in version 2.1.0 of the Python SDK). The SDK adds the required managed-agents-2026-04-01 beta header for you. The flow is three create calls and a stream:

from anthropic import Anthropic

client = Anthropic()  # reads ANTHROPIC_API_KEY

agent = client.beta.agents.create(
    name="Repo helper",
    model="claude-opus-5-5",
    system="You fix failing tests and explain what you changed.",
    tools=[{"type": "agent_toolset_20260401"}],
)

env = client.beta.environments.create(
    name="default-env",
    config={"type": "cloud",
            "networking": {"type": "limited", "allow_package_managers": True}},
)

session = client.beta.sessions.create(
    agent=agent.id,
    environment_id=env.id,
    budget={"type": "limit",
            "max_list_cost": {"amount": "500", "currency": "USD"}},  # $5.00 cap
)

You then open client.beta.sessions.events.stream(session.id), send a user.message event, and handle agent.message, agent.tool_use and session.status_idle events as they arrive. The quickstart has the full loop in eight languages. There is also an ant command line tool that creates agents from a Markdown file with front matter (ant apply) and can attach your terminal to a live session to approve tool calls.

Claude Managed Agents pricing

Billing has two parts, per Anthropic's pricing page:

Line item	Rate
Model tokens	The model's normal API rate (Claude Opus 5.5: $4 input, $20 output per million tokens)
Session runtime	$0.08 per session hour, metered to the millisecond while the status is running
Web search	$10 per 1,000 searches
US only inference	1.1 times the token rates when `inference_geo` is pinned to "us"

Runtime only accrues while a session is actually working. Time spent idle (waiting for your next message or for a tool approval), rescheduling or terminated is free, and runtime replaces the separate container hour charge that the code execution tool uses elsewhere. Prompt caching discounts apply as usual, but the Batch API discount does not, because sessions are interactive and have no batch mode.

Anthropic's own worked example is a one hour coding session on Opus 5.5 that uses 50,000 input tokens and 15,000 output tokens: $0.20 of input, $0.30 of output and $0.08 of runtime, for $0.58. With 40,000 of those input tokens served from cache, the total drops to about $0.43. The lesson from that example is that tokens, not runtime, usually dominate the bill. Even an agent that ran a full eight hours on every weekday of a month (about 176 hours) would add roughly $14 of runtime; the tokens it burned in that time would almost certainly cost more.

Keeping spend under control

Set a session budget. The amount is a string of whole US cents, so "500" means $5.00. The platform prices tokens, searches and runtime at list rates as the session runs, and once that total reaches the cap it stops starting new model requests. The request already in flight finishes, so the final figure can land slightly over. The session pauses rather than dying, and raising or removing the budget resumes it. A budget can only be added when a session is created, and deployments apply the same budget to every session they start.

Cloud sandbox or self hosted sandbox

By default each session gets a fresh Linux container on Anthropic's infrastructure with common languages and tools preinstalled, and you only choose network rules and extra packages. The alternative is a self hosted sandbox. Claude still runs on Anthropic's side and decides what to do next, but a worker process you run on your own machines claims each session from a queue, executes the bash and file tool calls locally and posts the results back. The ant CLI and the Python, TypeScript and Go SDKs ship ready made workers.

Anthropic's comparison page is blunt about what this does and does not change:

Files and processes stay on your worker, which suits data that cannot leave your network and internal services that are not publicly routable.
Conversation history, tool results, skills and memory stores still reach Anthropic, because the model has to see each result.
Web search and web fetch still run on Anthropic's side.
Compliance coverage and data retention do not change.

Sandbox providers have published worker guides, including E2B, Daytona, Modal, Cloudflare, Vercel, Fly.io and AWS Lambda. If you are choosing between those for agent workloads generally, our E2B vs Daytona vs Modal comparison covers the tradeoffs.

Limits and gotchas

Beta status. Every endpoint requires the beta header and Anthropic says behaviors may be refined between releases. MCP tunnels (for private MCP servers) and "dreaming" are a narrower research preview that needs an access request.
No ZDR or HIPAA. Because sessions are stateful and stored server side, the service is not eligible for Zero Data Retention or a HIPAA Business Associate Agreement. You can delete sessions and files through the API.
Rate limits. Per organization: 300 requests per minute on create endpoints and 1,200 on read endpoints, on top of normal usage tier limits.
Cloud availability. The pricing page says Managed Agents is not available on partner operated cloud platforms such as Amazon Bedrock. Anthropic documents two cloud routes where it does work: Claude Platform on AWS (billed in Claude Consumption Units) and, since October 22, Claude Models on Google Cloud (Anthropic managed), billed through Google Cloud Marketplace. Both have some feature differences. On Google Cloud, the prebuilt sandbox workers in the ant CLI and the SDKs do not work yet, so a self hosted sandbox there needs your own worker, per Anthropic's Google Cloud guide.
Permissions default to open. The built in toolset defaults to always_allow, while MCP tools default to always_ask. The newer auto policy has the server evaluate each call and allow, deny or pause it. Decide this deliberately before you point an agent at anything real.
Branding. Products built on it may say "Powered by Claude" but must not present themselves as Claude Code or Claude Cowork.

Claude Managed Agents alternatives

If you want the harness without the hosting, Anthropic's Agent SDK exposes the Claude Code loop as a Python or TypeScript library you run yourself, in your own sandbox, with any retention policy you like. You then own scaling, scheduling and state.

If you want a hosted style platform you can self host, the project people are comparing it with this week is Omnara, which describes itself as an open source managed agent platform and a self hostable alternative to Claude Managed Agents. It is Apache 2.0 licensed, runs agents on any model and machine through an API, and also sells a hosted cloud version. It is young, so test it before trusting it with production work.

If you would rather assemble the pieces, combine a framework such as LangGraph or the OpenAI Agents SDK with a sandbox provider and a memory layer like Letta or Mem0 (compared in our AI agent memory guide). That is the most work and the most control, and it keeps you free to switch models.

Who should use it

Pick Claude Managed Agents if you are already committed to Claude, want long running agents with sandboxes, memory and schedules without operating any of it, and do not need ZDR. Pick the Agent SDK or a framework if you need data retention control, other model vendors, or predictable infrastructure you can inspect end to end.

FAQ

How much does Claude Managed Agents cost?

You pay the model's normal token rates plus $0.08 per session hour while a session is running, and $10 per 1,000 web searches. Idle time is not billed. Anthropic's example one hour Opus 5.5 coding session costs $0.58 before caching.

Is Claude Managed Agents free to try?

There is no separate free tier listed. Access is enabled by default for every Claude API account, so you can try it with normal API credit and a small session budget as a safety cap.

Can Claude Managed Agents be self hosted?

Partly. You can run the sandbox, meaning where tool calls and files live, on your own infrastructure with a worker. The model and the agent loop always run on Anthropic's side, and conversation history and tool results still go to Anthropic.

Is Claude Managed Agents available on Bedrock?

No. Anthropic's pricing page says it is not available on partner operated cloud platforms. The documented AWS option is Claude Platform on AWS, where usage is invoiced in Claude Consumption Units and some features differ. Since October 22, 2026 it is also available on Claude Models on Google Cloud (Anthropic managed).

What is the difference between Claude Managed Agents and the Claude Agent SDK?

Both give you a Claude Code style agent loop. The Agent SDK is a library you run and host yourself. Managed Agents is the same idea as a hosted API, with Anthropic running the loop, sandboxes, session storage, schedules and memory.

Which models does it support?

You set the model on the agent, for example claude-opus-5-5, and self hosted sandboxes support every Claude model available in Managed Agents. Fast mode is available on supported Opus models at premium token rates.

Facts, prices and features in this guide were checked against official sources at the time of writing and can change. Always confirm details on the vendor's website.