Claude Managed Agents: Pricing, How It Works, and What Business Agents Cost (2026)

What is Claude Managed Agents?

Claude Managed Agents is Anthropic's hosted agent runtime: you define an agent (model, instructions, tools), and Anthropic runs it in a managed sandbox where Claude can read files, run commands, search the web and call your systems, for minutes or hours at a time. As of October 15, 2026 it costs standard Claude token rates plus $0.08 per session-hour, and the hour is only counted while the agent is actually running. It is in beta, it is switched on for every Claude API account, and it is not eligible for Zero Data Retention or a HIPAA Business Associate Agreement. For a business, that makes it a way to run back-office agents without building the agent loop and servers yourself, with one meter that matters more than people expect: not the runtime, but what the agent does while it runs.

How does Claude Managed Agents work?

Anthropic's overview page names four pieces. An agent is the model, system prompt, tools, MCP servers and skills, created once and reused by ID. An environment is where it runs: an Anthropic-managed cloud sandbox, or a self-hosted sandbox on your own infrastructure for compliance or data-residency reasons. A session is one running instance doing one task. Events are the messages between your application and the agent, streamed back as it works, with the history stored on Anthropic's side so a session can pause and resume.

The built-in tools are Bash, file operations (read, write, edit, search), web search and fetch (which you can restrict to an allowlist of domains), and MCP servers for your own systems. Two features matter for operations work: scheduled deployments start sessions on a cron schedule, down to the minute, so an agent can run every weeknight without anyone pressing a button, and vaults hold the credentials it uses to log in to your tools. Every request needs the managed-agents-2026-04-01 beta header, which Anthropic's SDK adds for you.

How much does Claude Managed Agents cost?

Two meters, per Anthropic's pricing documentation and claude.com/pricing, both read October 15, 2026:

• Tokens, at the same rates as the Claude API, with the same prompt-caching discounts.
• Session runtime: $0.08 per session-hour, measured to the millisecond and charged only while the session's status is running. Time spent idle (waiting for your next message or a tool approval), rescheduling, or terminated is free. Runtime replaces code-execution container billing, so there is no separate container-hour charge on top.

Three extras apply inside a session: web search is $10 per 1,000 searches, pinning inference to the US costs 1.1x on tokens, and fast mode on Opus 5.5 is billed at 2x. One discount does not apply: the Batch API's 50% off, because sessions are stateful and interactive. Anthropic's own worked example is a one-hour coding session on Opus 5.5 with 50,000 input and 15,000 output tokens: $0.20 + $0.30 + $0.08 of runtime = $0.58, or $0.428 when 40,000 of the input tokens are cache reads.

What three business agents would cost per month

Anthropic's example is a coding session. Here is the same arithmetic on work a service business would actually hand an agent. The volumes are our assumptions; the rates are the published ones above.

The finding: runtime is the smallest meter. In all five rows the $0.08 hour is under 10% of the bill (about 5% for the reconciliation agent, 9% for triage). What moves the number is the model you choose, how much of the prompt is cached, and how many web searches the agent fires. On a cheap model, searches can be most of the cost, so an allowlist of the three sites your agent actually needs is a pricing decision, not just a safety one. And because idle time is free, an agent that stops to wait two days for a manager's approval costs nothing while it waits. If the bill still grows, the levers are the same as on any Claude system, covered in our guide to AI code optimization.

Claude Managed Agents vs building on the Messages API

Anthropic positions the two as different tools. The Messages API is direct model access for teams that want their own agent loop and fine control. Managed Agents is the pre-built harness for long-running and asynchronous work: multi-step tasks with many tool calls, a sandbox with packages and network access already set up, persistent files and history between sessions, and runs on a schedule.

• Pick Managed Agents when the job runs for minutes, uses tools, and nobody on your team wants to own servers, sandboxes or retry logic: the nightly reconciliation, the weekly report, the research run before every sales call.
• Pick the Messages API when the job is a single pass (classify this email, extract these invoice fields, summarize this call) and can wait. That is where the 50% Batch discount lives, and it does not exist inside Managed Agents.
• Pick neither for a live phone call. Managed Agents is built for asynchronous work; a caller on the line needs a voice platform built for real-time conversation.

If you would rather not touch an API at all, the no-code platforms wrap the same models in a visual builder at a per-seat or per-credit price; we priced nine of them on one workload in best no-code AI agent builders.

What Claude Managed Agents does not cover

• Regulated data. Because sessions store history, files and sandbox state on Anthropic's side, Managed Agents is not eligible for Zero Data Retention or a HIPAA BAA. A medical practice or anyone handling protected health information should not run patient data through it. You can delete sessions and uploaded files through the API.
• Stability guarantees. It is a beta, and Anthropic says behaviors may be refined between releases. MCP tunnels and the "dreaming" memory feature are in a more limited research preview that needs an access request.
• Other clouds. It is available on Claude Platform on AWS, with some feature differences, but not on the other partner-operated cloud platforms.
• The build. Anthropic hosts the runtime. Deciding what the agent may do on its own, connecting it to your CRM or accounting system, and testing it on your real data is still the work, and that is where most of the cost of an agent sits; the market ranges are in how much an AI agent costs.

Who can build a Claude agent for your business?

Your own developer can, if you have one who wants to own it. Otherwise it is an agency job, and the decision usually comes down to the points in our build vs buy AI agents guide. SuperDupr, the AI ops shop Justin McKelvey runs, builds agents on Claude and other models for service businesses at published prices as of October 2026: $5,000 to $10,000 for a single-purpose agent, $8,000 to $25,000 for larger intake, follow-up and operations agents, and $25,000 to $50,000+ for multi-agent systems, plus $200 to $2,000 a month for hosting, model usage and optimization. Most go live in 2 to 4 weeks and the business owns the system. The details are on our AI agent development page, and the first step is a free 30-minute strategy session.

The bottom line

Claude Managed Agents turns "build an agent and run it somewhere" into an API call, priced at normal Claude token rates plus $0.08 for each hour the agent is actually working. For back-office jobs the runtime charge is close to a rounding error; the model, the cache and the web searches set the bill, so price your agent on one real run before you scale it. Keep regulated data out of it while it lacks ZDR and BAA coverage, and use the Messages API Batch discount for single-pass work that can wait.

Sources: platform.claude.com/docs Managed Agents overview, scheduled deployments, and the pricing page's "Claude Managed Agents pricing" section; claude.com/pricing (model and feature rates). All read October 15, 2026. Monthly figures in the worked table are our arithmetic on assumed workloads.