# Claude Managed Agents

Official pricing: https://platform.claude.com/docs/en/about-claude/pricing#claude-managed-agents-pricing  
Category: agent-sandbox · Isolation: container

## Pricing regimes (raw)

- **Cloud sandbox session runtime ($0.08 per running session-hour, flat)** (sizes): cloud sandbox (fixed) ? vCPU/8 GiB $0.08/h [alt]
- **Self-hosted sandbox environment (your infra, Anthropic orchestration)** (sizes): self-hosted worker ? vCPU/? GiB $null/h [alt]
- **Messages API code execution tool container (not Managed Agents)** (sizes): code execution container 1 vCPU/5 GiB $0.05/h

## Features

Yes: snapshots, pause/resume, persistent disk, volumes, idle auto-stop, ≥24 h sessions, custom image (Docker or snapshot), start from your own snapshot, your own Docker/OCI image, browser, preinstalled agents, egress allowlist, open internet, self-hosting, BYOC, SOC 2, SSO, Python, Node.js, Go, Rust, Java, credential injection, wake on request, webhooks, scheduled runs (cron), spend limits, MCP server, automatic snapshots, agent harness API, secret proxy, secret proxy for any API, extra volumes, shared volumes

No: memory snapshots, fork/clone, full VM (own kernel), nested virtualization, desktop GUI, computer-use API, code interpreter, browser + desktop control, anti-bot stealth, CAPTCHA solving, residential IPs, SSH, public IPv4, HTTPS preview URLs, custom domains, raw TCP inbound, static egress IP, open source, HIPAA, GPU, arm64, Windows, macOS, GPU desktop, HTTP method/path egress rules, live resize, live fork (no pause), memory fork, EU data residency, zero data retention, snapshots on demand, hosted agent API (their own agent), gVisor or VM (no shared kernel), inbound access rules, firewall inside (nftables)

Unknown: everything else. Evidence (source + quote) per feature: https://battleships.dev/data/providers/claude-managed-agents.json → feature_evidence

## Caveats

- Managed Agents session runtime (on-demand) is flagged alt / product_class agent-platform: it only exists inside a Claude agent session and LLM tokens are billed on top (not modeled). Sandbox vCPU unpublished -> shown as unsized.
- Code execution tool containers have NO outbound network access (pre-installed libraries only).
- Not a general-purpose sandbox: the runtime only exists inside a Claude agent session, and model tokens (not modeled by this card) usually dominate the bill. In Anthropic's own example a 1 h Opus 5 session costs $0.625 in tokens vs $0.08 runtime.
- Flat $0.08/h regardless of resources; the sandbox vCPU count is unpublished (size vcpu null), memory 'up to 8 GB', disk 'up to 10 GB'. Workloads needing >8 GiB RAM are infeasible.
- 'running' includes time the agent spends waiting on model inference, not only CPU activity; cpu util does not reduce cost.
- Idle sessions retain their sandbox free of charge; retention limit for idle sessions/sandboxes is unpublished. Deleting a session deletes its sandbox and session-produced files.
- Self-hosted environment runtime billing not documented (mode hour null, flagged alt).
- Code execution tool (Messages API) is a separate SKU: $0.05/container-hour after 1,550 free hours/org/month, 5-min minimum, 1 CPU/5 GiB, no internet; free when used with web_search/web_fetch 20260209+. The 1,550 free hours ($77.50) are encoded as included_usd on the API plan, scoped to the code-execution-tool mode.
- Egress/IPv4 not priced or offered (no inbound ports). Region of cloud sandboxes unpublished (regions []).
- Beta since 2026-04-08; runtime price unchanged at $0.08 since launch. Some third-party blogs quote $0.25/session-hour — not supported by the official pricing page.
- Not eligible for HIPAA BAA or ZDR. No concurrency cap published; creates limited to 300/min per org.
- No benchmark data in benchmarks.json.
- Docker inside the cloud sandbox is listed as 'limited availability' (keep features.docker_inside null).
- Other limits: max 1,000 scheduled deployments per org (support can raise); max 500 files per session; max 8 memory stores per session; memory versions retained 30 days. Dreams and advisor consultations are billed at token rates only.
- Multiagent sessions: up to 25 concurrent threads (coordinator + subagents) share ONE sandbox and filesystem, and runtime is metered per session status, so parallel subagents do not multiply the $0.08/h runtime (they do multiply tokens).

## How this provider charges


Anthropic's hosted agent harness (public beta since 2026-04-08) bills on **two dimensions: tokens + session runtime**.
The sandbox part is a single flat SKU: **$0.08 per session-hour**, metered to the millisecond, **only while the
session status is `running`**. There is no size choice: each session gets an Ubuntu 24.04 container with up to 8 GB RAM
and 10 GB disk (vCPU count unpublished). Tokens are billed at normal API list rates and usually dominate.

## Regime table

| Regime | When it applies | How billed | Numbers | Source |
|---|---|---|---|---|
| Session runtime – running | Agent actively executing (tool calls **and** waiting on model inference) | Flat per session-hour, ms metering; resource-independent | $0.08/h | https://platform.claude.com/docs/en/about-claude/pricing |
| Session idle | Waiting for a user message / tool confirmation; also after a budget cap is hit | Not billed; sandbox + history preserved | $0 | pricing page, /managed-agents/session-operations, /managed-agents/budgets |
| Rescheduling | Transient error, auto-retry | Not billed | $0 | pricing page |
| Terminated / archived | Unrecoverable error or archived | Not billed; history kept | $0 | same |
| Deleted | `DELETE /sessions/{id}` | Removes record, events, sandbox and session-produced files | $0 | /managed-agents/session-operations |
| Tokens | Every model request in the session | Model list rates; prompt-caching multipliers apply; **no Batch discount**; fast-mode premium if `model.speed=fast` | e.g. Opus 5 $5/$25, Sonnet 5.5 $2/$10, Haiku 4.5 $1/$5 per MTok | pricing page |
| US-only inference | `model.inference_geo = "us"` | 1.1× on tokens only | ×1.1 tokens | pricing page |
| Web search in session | Each search | Per search | $10 / 1,000; web fetch free | pricing page |
| Self-hosted sandbox environment | `type: self_hosted`, worker on your Linux host | You pay your compute; whether the $0.08/h runtime applies is **not documented** | null | /managed-agents/self-hosted-sandboxes |
| Claude Platform on AWS | Using MA via AWS Marketplace | Same token + runtime charges converted to CCUs at $0.01; autonomous sessions need re-auth after 6 h | same rates | pricing page, /build-with-claude/claude-platform-on-aws |
| Session budget | Optional per-session hard cap | Enforced at **list** cost (tokens + searches + running time); session pauses idle at the cap, overshoot ≤ one model request per thread | user-set | /managed-agents/budgets |
| Enterprise / volume | Negotiated | Custom rate limits, volume discounts | unpublished | pricing page |
| Code execution tool (separate SKU, Messages API) | `code_execution` tool outside Managed Agents | Execution time per container, 5-min minimum; 1,550 free h/org/month; **free** with web_search/web_fetch 20260209+; replaced by session runtime inside MA | $0.05/container-h; 1 CPU / 5 GiB / 5 GiB, no internet, 30-day expiry | pricing page, /agents-and-tools/tool-use/code-execution-tool |
| Storage / egress | Session filesystem, outbound traffic | No charge published | null / not priced | — |

No dated runtime price change since launch (the $0.08 figure was quoted at the 2026-04-08 launch and is unchanged).
Token prices did change: Sonnet 5's $2/$10 intro price became permanent (the planned rise to $3/$15 on 2026-09-01 was cancelled).

## Gotchas

1. **"Running" ≠ CPU busy.** Time the agent waits on model inference counts as running. Only time waiting on *you* is free.
2. **No size choice.** vCPU is unpublished and RAM tops out at 8 GB. The same $0.08/h is charged whether the agent uses 1 core or 4.
3. **Idle sessions are a free pause:** the sandbox is kept, with no runtime charge and no published retention limit.
4. **Tokens are the bill.** In Anthropic's own worked example (1 h, Opus 5), tokens are $0.625 and runtime is $0.08, so runtime is only 11% of the total.
5. **Budgets use list prices**, so negotiated discounts don't stretch a budget.
6. Not eligible for HIPAA BAA or ZDR. No inbound ports, SSH or IPv4. Egress is allowlist-capable (`limited` networking).
7. Third-party posts quoting "$0.25/session-hour" are wrong per the official pricing page.

## Worked example

4 vCPU / 8 GiB, 50 concurrent × 8 h/day × 22 days = 8,800 session-hours, 30% CPU, 50 GiB snapshots, 100 GiB egress.
The shape is taken as the fixed sandbox (8 GB RAM fits; vCPU unknown). Assume the agent is `running` the whole 8 h.

| Line | Calculation | Monthly |
|---|---|---|
| Session runtime | 8,800 × $0.08 | **$704.00** |
| CPU util 30% | no effect (flat SKU) | $0 |
| Snapshots 50 GiB | idle-session state not billed | $0 |
| Egress 100 GiB | not priced | $0 (unpublished) |
| **Sandbox subtotal** | | **$704.00** |
| Illustrative tokens | Anthropic's example rate of $0.625 per running hour on Opus 5, × 8,800 | +$5,500 |
| If agent is running only 50% of wall-clock (rest idle awaiting input) | 4,400 × $0.08 | $352 runtime |

Feasibility: no published concurrency cap; 50 sessions/day is far under the 300 creates/min limit.