Claude Managed Agents Pricing: What It Costs in 2026

Do not index

Claude Managed Agents pricing has two parts: standard Claude API token rates, plus $0.08 per session-hour of runtime. There is no monthly fee, no per-agent license, and no plan requirement; any Claude API key works. Runtime is metered to the millisecond and only accrues while a session is actually running, so a genuinely 24/7 agent costs about $58 a month before a single token, while a mostly-idle one costs far less. Here is the full September 2026 breakdown, the fine print the April launch posts still get wrong, and what the flat-price alternatives cost.

Claude Managed Agents pricing at a glance

Anthropic launched Claude Managed Agents in public beta on April 8, 2026. It is an API product: you define an agent (model, prompt, tools, skills), Anthropic runs it in their own harness and sandbox, and you pay for what it consumes. The official pricing page bills it on two dimensions, tokens and session runtime, plus standard per-use tool charges like web search:

Meter	Rate
Session runtime	$0.08 per session-hour, billed to the millisecond
Model tokens	Standard Claude API rates (see below)
Web search inside sessions	$10 per 1,000 searches

Token rates are the same as the regular API. Per million tokens (input/output): Haiku 4.5 is $1/$5, Sonnet 5 is $2/$10, Opus 5 is $5/$25. Sonnet 5's $2/$10 rate is now the standard price; Anthropic cancelled the increase to $3/$15 that was scheduled for September 1, so any pricing guide still warning about it is out of date.

Two things do NOT apply inside Managed Agents sessions: the Batch API discount (sessions have no batch mode) and partner-cloud pricing from Bedrock or Vertex. Two things DO apply that the launch-week posts never mention: a fast-mode premium when an agent's model.speed is set to "fast", and a 1.1x token multiplier when model.inference_geo is pinned to "us".

What counts as billable runtime

Runtime only accrues while the session's status is running. Idle, rescheduling, and terminated time is free. That single rule is the difference between a cheap bill and an expensive one, and it is worth reading twice.

Anthropic's own worked example: a one-hour Opus 5 coding session consuming 50,000 input and 15,000 output tokens costs $0.705 total. That is $0.25 of input, $0.375 of output, and $0.08 of runtime. With prompt caching covering 40,000 of those input tokens, the same session drops to about $0.53. Notice the ratio: even in Anthropic's example, tokens are almost 8x the runtime fee.

Session runtime also replaced the old code-execution container-hour billing. You are not charged container hours on top of session-hours; it is one meter now.

The always-on math: what a 24/7 agent really costs

Here is the calculation none of the ranking pricing guides do. An average month has about 730 hours, so an agent whose session runs around the clock costs 730 x $0.08 = $58.40 in runtime alone ($57.60 in a 30-day month). Tokens come on top.

And tokens dominate. A continuous workload that burns 2 million input and 500,000 output tokens a day on Sonnet 5 adds $9 a day ($4 input + $5 output), roughly $270 a month, nearly five times the runtime fee. On Opus 5, the same volume is about $675 a month. The $0.08 meter is not what empties the account; the model is.

The distinction that matters is always-available versus always-running. If your agent wakes on a webhook or a cron schedule, does its work, and goes idle, you pay only for the running minutes, and Managed Agents is genuinely cheap. If it needs to hold a continuous loop (a monitoring agent, a queue worker, an agent that lives in a chat channel and must react instantly), you are paying the full ~$58 runtime floor plus a heavy token bill, every month, per agent.

The fine print the launch posts miss

Most pages ranking for this query were written in April 2026 and never updated. What has changed or was never covered:

Session budgets. Since August you can set a hard spend cap per session; when it hits, the session pauses with a budget_reached stop reason instead of burning more money. This is Anthropic's answer to the runaway-bill fear that dominated the launch discussion.

The new tokenizer. Models from Claude 4.7 onward tokenize the same text into roughly 30% more tokens, so per-token prices understate cost growth if you are comparing against early-2026 estimates.

US data residency costs extra. Pinning inference to US infrastructure bills tokens at 1.1x.

No Managed Agents-specific trial. There is no Managed Agents credit or trial tier; beyond the small free credit Anthropic grants new API accounts, billing starts with the first session-hour on your API key.

Still in beta. Every endpoint requires the managed-agents-2026-04-01 beta header. It is enabled by default for all API accounts, and no general-availability date has been announced.

What you get for the money, and the limits

The $0.08 buys real infrastructure: Anthropic's managed harness, sandboxed environments (you can self-host the sandbox if you want your own compute), persistent memory, multiagent orchestration, webhooks, scheduled cron deployments, and, since September 10, attaching to a live session from your terminal. You run zero servers.

The limits are just as concrete. The harness is Anthropic's: if you want your own agent loop, you are back to the Messages API. Sessions store state server-side, which is why Managed Agents is not eligible for Zero Data Retention or a HIPAA BAA. There are no built-in messaging channels; input and output are API events, SSE streams, and webhooks, so wiring the agent into Telegram or WhatsApp is your integration work. MCP tunnels and the "dreams" memory-curation feature (Claude reflects on past sessions to clean up an agent's memory store) are request-access previews. And org-wide rate limits sit at 300 create-requests and 1,200 read-requests per minute.

Managed Agents vs Claude Code on the web vs Cowork

Three Anthropic products get mixed up in every pricing thread, but they sit on only two billing models: one bills by usage, the other two are bundled into a Claude subscription.

Claude Managed Agents is the API product priced above: tokens plus session-hours, for developers building agents into their own software.

Claude Code cloud sessions are in research preview for Pro, Max, and Team plans (and Enterprise premium or Chat + Claude Code seats) at no separate compute charge, but the VMs are task-scoped and expire when idle. They are for kicking off coding tasks, not for keeping an agent alive. We cover the always-on gap in our Claude Code cloud guide.

Claude Cowork is the consumer agent inside the Claude app. It has no separate price line, and in mid-September 2026 Anthropic announced it is folding its capabilities into the main Claude chat interface, rolling out to Pro and Max plans first. If you were evaluating it for work Claude's app can't hold, see our Cowork alternatives comparison.

When per-hour billing wins, and when flat-fee wins

Usage billing is the right shape for bursty work. An agent that runs 20 hours a month costs $1.60 in runtime; no flat-fee box beats that, and idle-is-free means you never pay for a weekend of silence. If your workload scales to zero, Managed Agents is priced honestly for it.

Flat fees win the moment the agent is genuinely always on. Here is the always-on month, September 2026 prices, verified against each provider's live pricing page:

Option	Billing shape	Always-on month, before model usage
Claude Managed Agents	$0.08/session-hour + tokens	~$58
DigitalOcean 2 vCPU/4 GB droplet (DIY)	$24/mo flat	$24
Nous Hermes Cloud	$0.56/day while running	~$17
E2B / Daytona sandboxes	per-second; E2B caps sessions at 1-24 h by plan	not built for always-on
Agent37 dedicated instance	$4.94/mo flat	$4.94

A note on the DIY row: the EUR 5.49 Hetzner CX23 that older guides recommend is effectively not orderable right now (Hetzner's cloud page marks the whole Cost-Optimized line "currently not available"), so the realistic DIY floor is a $24 droplet or Hetzner's CPX line from EUR 19.49. DIY also costs you the ops: provisioning, patching, restarting the agent when it falls over. The sandbox platforms are built for ephemeral code execution: E2B caps sessions at 1 hour (Hobby) to 24 hours (Pro), and an always-on 2 vCPU / 4 GB Daytona sandbox runs roughly $120 a month at its per-second rates. A persistent agent is the wrong workload for them.

The flat-price alternative: run the harness yourself on Agent37

Here is the trade we sell, stated plainly. Claude Managed Agents gives you Anthropic's harness with zero ops and a usage meter. Agent37 Cloud gives you your choice of open-source harness on a persistent instance with a flat bill: a dedicated 2 vCPU / 4 GB instance is $4.94 a month, running 24/7, and shared capacity starts at $1.99 a month (enabled per workspace).

One API call provisions an always-on agent computer from a template: POST /v1/instances. There are eight system templates, including Claude Code, Codex, OpenCode, OpenClaw, the Hermes agent, and n8n, and any public Docker image works as a custom template. The agent37-claude-code template runs on your own Anthropic account (sign in with claude auth login, an OAuth token, or an API key), so model usage bills at exactly Anthropic's rates. There is no session-hour meter and no markup on tokens: model usage bills your own Anthropic account directly, and Agent37 never supplies, meters, or bills it.

Everything persists until you delete the instance: files, memory, sessions, connected tools. Scheduled cron runs, webhooks, public ports, and custom domains are built in, and the whole thing is white-label by default if you are reselling agents to your own customers. Agent37 is SOC 2 Type I audited (Type II in progress), backed by Y Combinator, and runs production agents for customers like Boon and ApolloClaw. The honest downside: at these prices support is an AI bot first, not a human on call.

If your agent needs to exist all day rather than run in bursts, the math is not close: $4.94 flat versus ~$58 plus tokens. Deploy one with a single API call, always-on from $4.94/mo dedicated or $1.99/mo shared, with a $1 starter credit and no card required, at agent37.com/cloud.

FAQ

Is there a monthly fee for Claude Managed Agents?

No. Pricing is pure usage: standard Claude API token rates plus $0.08 per session-hour while the session is running. Any API account has it enabled by default.

Did Claude Managed Agents pricing change since launch?

The base rate has not changed: $0.08 per session-hour since the April 8, 2026 launch. Posts quoting $0.25 per session-hour contradict Anthropic's live pricing page. What launch coverage missed are the additions: session budgets and the 1.1x US inference multiplier (both added August 7, 2026), plus a fast-mode premium.

Does the Batch API discount apply to Managed Agents?

No. Sessions have no batch mode, so all tokens bill at standard rates. Prompt caching does work, and cache reads bill at a tenth of the input rate.

Is there a free trial for Claude Managed Agents?

There is no Managed Agents-specific credit or trial. New API accounts get a small amount of free general credits; beyond that, you pay from the first session-hour on your API key.

What is the cheapest way to keep a Claude-based agent running 24/7?

Run the harness on your own account on a flat-fee persistent box. An Agent37 dedicated instance with the Claude Code template is $4.94 a month always-on, with model usage billed by Anthropic to your own account at their normal rates. A DIY VPS (a $24 DigitalOcean droplet, or Hetzner's CPX line from EUR 19.49 while the cheap CX line is unorderable) is the same idea with the ops on you.

Claude Managed Agents Pricing: What It Actually Costs in 2026