Claude Managed Agents Pricing
What a session costs, when the runtime fee matters, and how other agent platforms price it
On this page
Claude Managed Agents bills for two things: the tokens a session uses, at Anthropic's normal API rates, and $0.08 for every hour the session is running. Time a session spends idle is free. In Anthropic's own example, an hour on Claude Opus 5.5 costs $0.58, and $0.50 of that is tokens.
The runtime fee is small next to frontier-model tokens and large next to small-model tokens. So the model you pick, and how many hours your agents run, decide whether it matters.
The Short Version
How the Bill Works
Tokens. Every token in a session is billed at the standard model rates: $4 per million input tokens and $20 per million output tokens on Claude Opus 5.5, $2 and $10 on Claude Sonnet 5.5, and $0.10 and $0.50 on Claude Haiku 5.5. Prompt caching applies the same way it does on the Messages API, and a cache hit on Opus 5.5 or Sonnet 5.5 costs 5% of the input price.
A few modifiers change the token bill. Pinning an agent to US-only inference multiplies token prices by 1.1, fast mode adds its premium, and web search costs $10 per 1,000 searches. The Batch API discount doesn't apply, because sessions are interactive, and Managed Agents isn't available on partner-operated clouds such as Amazon Bedrock. On Claude Platform on AWS, session charges convert to Claude Consumption Units.
Runtime. Runtime costs $0.08 per session-hour, metered to the millisecond, and it only accrues while the session's status is running. Time spent waiting for your next message or for a tool confirmation counts as idle and costs nothing. The runtime fee also covers the session's container, so there's no separate container-hour charge.
Your own sandbox. You can run a session's commands on your own machines with self-hosted sandboxes. Anthropic's pricing meters runtime by the session's status and lists no exception for them, so a session on your machines still accrues $0.08 an hour, and you pay for the machines as well.
What Drives the Cost
The model. Anthropic's example hour uses 50,000 input tokens and 15,000 output tokens. At that mix, the runtime fee is 14% of the hour on Opus 5.5 and about a quarter on Sonnet 5.5. On Haiku 5.5, the same tokens cost about a cent, so the $0.08 runtime fee is 86% of the bill.
Hours, at fleet scale. An agent that runs eight hours every workday accrues about 176 session-hours a month, or $14 in runtime. A hundred of them cost $1,408 a month in runtime before any tokens. On Haiku 5.5 with the same token mix, their tokens come to about $220, so the runtime fee is most of what you pay.
What the agent waits on. A session waiting on a person or on your app is idle, so approvals that take a day cost nothing in runtime. A long build or test run is different, because the session is running the whole time.
Caching. In Anthropic's example, caching 40,000 of the 50,000 input tokens brings the Opus hour from $0.58 to $0.43. Agents that resend the same instructions and files every turn save the most.
Small models can need more tries to finish a task, so compare what a finished task costs, not just an hour. To cap spend, a session budget pauses a session between model calls once it reaches a limit. It covers one session, so limits per customer are yours to build.
How Other Platforms Price Agents
Every platform below bills model tokens at its provider's API rates. The difference is what you pay for the time an agent runs, and for the time it waits:
The per-hour fees look small, and on a frontier model they are. They decide the bill for fleets of small-model agents that run all day, and for agents that keep a container busy with builds and tests.
Omnara
Omnara is an open-source managed agent platform with no platform fee and no charge per session-hour. With your own model keys and machines, Omnara charges nothing, and you pay only your model provider. An agent that only calls your APIs through custom tools doesn't need a machine at all.
That changes the fleet example above. A hundred Haiku 5.5 agents that run eight hours a workday and call your APIs through custom tools cost about $220 a month in tokens on your own Anthropic key, with nothing added for the hours they run.
Omnara works with models from OpenAI, Anthropic, OpenRouter, Amazon Bedrock, or any compatible endpoint, and each subagent can use a different one. A small model can do the routine steps while Opus handles the hard ones, and you can change a running agent's model through the API. Omnara's own loop also spends fewer tokens. On Terminal-Bench 4.0, it ran 5% cheaper than Claude Code on Claude Opus 5.5, and 11% cheaper than Codex on GPT-6.1 Sol with 35% fewer tokens (traces).
On Omnara Cloud, models bought through Omnara cost the provider's token rates, and Omnara's machines are billed by usage and sleep when idle. Both are paid with credits that carry a 5% service fee covering provider costs. The API reports model usage and provider-reported cost by organization, project, and agent, so your app can enforce limits per customer. Self-hosting is free under Apache 2.0.
Which One to Choose
Claude Managed Agents fits Claude-only agents where most of the spend is Opus or Sonnet tokens and you want nothing to host. At Anthropic's example mix, the runtime fee is 14% to 24% of each hour.
For fleets of agents that run for hours, agents on small models, or agents on more than one lab's models, choose Omnara. It has no platform or session-hour fee, your own keys and machines cost nothing, and its pricing is published.
FAQ
How much does Claude Managed Agents cost?
You pay Anthropic's standard token rates plus $0.08 for each hour a session is running. In Anthropic's example, an hour on Claude Opus 5.5 with 50,000 input and 15,000 output tokens costs $0.58: $0.50 in tokens and $0.08 in runtime.
Is idle time billed?
No. Runtime accrues only while a session is running. Waiting for your next message or for a tool confirmation counts as idle, so an approval that waits overnight adds nothing to the runtime fee.
Does running the sandbox on my own machines lower the price?
Not the runtime fee. Anthropic meters runtime by the session's status, so a session with a self-hosted sandbox still accrues $0.08 an hour, and you pay for your machines too.
Can I pay for Claude Managed Agents through AWS or Google Cloud?
Managed Agents isn't available on partner-operated clouds such as Amazon Bedrock. Through Claude Platform on AWS, session token and runtime charges convert to Claude Consumption Units at the standard rate.
What's the cheapest way to run Claude agents in production?
Run them on Omnara with your own Anthropic key. Omnara has no platform or session-hour fee, so you pay only Anthropic's token rates and whatever machines your agents use. You can also self-host Omnara for free.