Claude Managed Agents pricing

What you are billed for

Claude Managed Agents bills on 2 lines: tokens and session runtime. Tokens are charged at the normal Claude API model rates, and prompt caching discounts apply the same way they do on the Messages API.

Runtime costs $0.08 per session-hour. It is metered to the millisecond, and it only runs while the session status is running. Time the session spends idle, waiting for the next user message or for a tool confirmation, is free. So is time spent rescheduling or after the session is terminated.

Runtime replaces the container-hour billing of the code execution tool. You do not pay for container hours on top of it. Web search inside a session is the only common extra: $10 per 1,000 searches, the same as on the API.

Model rates that drive the bill

Per 1 million tokens, input and output, from Anthropic's pricing page on 2 November 2026: Claude Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens, Claude Sonnet 5.5 at $2 and $10, Claude Opus 5.5 at $4 and $20, and Claude Fable 5.1 at $10 and $50.

Cache reads cost 5% of the input price on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1. An agent that rereads the same instructions, tools and documents every turn gets most of its input from cache, which is the single biggest saving available.

• US-only inference costs 10% more. If the agent pins inference to the US, every token line is billed at 1.1 times the standard rate.
• Fast mode is a premium tier. On Opus 5.5, fast mode is $8 input and $40 output per 1 million tokens, double the standard rate.
• No batch discount. Sessions are stateful and interactive, so the 50% Batch API discount does not apply.
• Not on partner clouds. Managed Agents is not available through partner-operated platforms such as Amazon Bedrock. On Claude Platform on AWS, the same charges convert to consumption units.

3 worked examples, per month

Each example uses list prices with no caching unless stated. Runtime assumes the session is actively running for the minutes shown, not the wall clock time a user keeps a chat open.

• Support agent on Sonnet 5.5: about $78 a month. 1,000 sessions, each with 20,000 input tokens, 3,000 output tokens and 6 minutes of running time. Per session that is $0.04 of input, $0.03 of output and $0.008 of runtime, so $0.078. Runtime is about 10% of the bill.
• Research agent on Opus 5.5 with search: about $720 a month. 500 sessions, each with 200,000 input tokens, 20,000 output tokens, 30 minutes of running time and 20 web searches. Per session: $0.80 input, $0.40 output, $0.04 runtime and $0.20 search, so $1.44.
• The same research agent with caching: about $435 a month. If 150,000 of the 200,000 input tokens are cache reads, input drops from $0.80 to $0.23 per session and the session total falls to $0.87. Same work, 40% less.

How to keep the bill under control

1. Set a session budget when you create each session. It is a hard cap in US cents, priced at public list rates across tokens, searches and runtime.
2. Leave margin above the cap. The check runs before each model request, so the request in flight finishes and a 50 cent cap can stop at 53 cents.
3. Handle the budget_reached stop reason. The session pauses and stays idle with its history and sandbox intact. Raising or removing the budget resumes it.
4. Decide budgets up front. A budget can only be attached when the session is created, and once removed it cannot be added back to that session.
5. Use deployment budgets for scheduled runs. The cap is copied onto each session the deployment starts, so it limits each run, not the total.
6. Pick the smallest model that passes your evaluation set, then turn on caching for the system prompt, tools and shared documents.
7. Read usage.list_cost and usage.active_seconds from each session and alert on outliers before the invoice does.

What we build around Managed Agents, and what it costs

• Add 1 agent to your live product, from $5,000, about 3 weeks. One agent or workflow wired into your app, with an evaluation set, session budgets, usage alerts and monitoring. Right when the product exists and the agent is 1 job.
• AI-native product build, $15,000 to $35,000, 6 to 10 weeks. A new product with agents designed into the architecture from week 1: data model, permissions, human approval steps and cost guards. Right when the agent is the product.
• Your own agent loop on the Messages API, scoped per project. When you need data handling that Managed Agents does not offer, batch pricing, or a partner cloud, we build the loop, tools and state in your own backend.

When Managed Agents is the wrong choice

Managed Agents is not eligible for zero data retention or HIPAA, because sessions are stateful and transcripts persist until you delete them. If the agent will handle health records or data your contracts say must not be stored, do not put it on Managed Agents. The same goes for large offline jobs, where the Batch API halves the token bill and Managed Agents cannot.