Claude Managed Agents Pricing: Runtime + Tokens | OpenMA

Claude Managed Agents (CMA) pricing has two moving parts: the tokens your agent consumes and the session runtime it spends in a running state. Both are usage-based and billed separately from Claude Code subscription plans.

Figures below come from Anthropic’s public pricing documentation as of November 2026. Rates change, so check that page before you budget. If you need a refresher on what the product is, start with what is Claude Managed Agents.

What Anthropic bills

Tokens

Every model call your agent makes is billed at the standard Claude API token rates for the model the agent uses: input, output, and prompt caching reads and writes, as listed on the pricing page. Managed Agents does not add a separate per-token markup. The harness makes more calls than a single chat request would, because an agent loops: read context, call the model, run a tool, feed the result back, repeat.

Session runtime

On top of tokens, Anthropic charges $0.08 per session-hour for session runtime. Runtime is metered while a session is in the running status — when the harness is actively working. A session sitting idle, waiting for your next message, is not accruing runtime.

That makes the runtime fee easy to reason about:

Running time per session	Runtime cost per session
6 minutes	$0.008
30 minutes	$0.04
1 hour	$0.08
8 hours	$0.64
Running 24/7 for a month (~730 h)	~$58.40

Not the same as Claude Code pricing

The Pro, Max, Team, and Enterprise subscription plans are seat-based plans for people using Claude’s apps and Claude Code. Claude Managed Agents is a developer platform API: it is billed by usage (tokens plus session runtime) through your Claude Platform account, separately from subscription seats.

If your team uses both, budget them separately.

Hidden and easy-to-miss spend

Tools and external calls

Server-side tools carry their own charges. Web search inside a session costs $10 per 1,000 searches, and every tool result you feed back into context becomes input tokens on the next turn. When a tool returns a 40 KB page, that page is billed as input tokens even if the tool call itself is free.

MCP servers and external APIs your agent calls are billed by whoever runs them. They do not appear on your Anthropic invoice, but they are part of the cost of a run.

Pricing modifiers

The pricing page lists a few modifiers that change the token line:

• Setting an agent’s model.speed to "fast" bills fast-mode premium rates.
• Pinning model.inference_geo to "us" bills tokens at 1.1x.
• The Batch API’s 50% discount does not apply, because sessions are stateful and interactive.

Session runtime replaces the code execution container-hour billing, so you are not charged container hours on top of it.

Retries and failed runs

A run that fails at step 18 of 20 has already paid for 18 steps of tokens and runtime. Tool errors that the agent retries, sessions you re-run after a bad outcome, and outcome-evaluation loops that iterate until a rubric is satisfied all multiply the base cost. Track cost per completed task.

Idle vs running session status

Runtime only accrues while running, which rewards agents that finish and go idle. It penalizes agents that poll, wait on slow tools while still “running,” or run long background jobs inside a session. Context growth is the other slow leak: long sessions re-send more history on every model call unless you manage compaction and caching well.

Worked monthly example

A support-triage agent, modeled for one month. Replace these with your own measurements, and the same formula applies.

Input	Value
Sessions per month	2,000
Average running time per session	12 minutes (0.2 h)
Average input tokens per session (all turns, after caching)	150,000
Average output tokens per session	10,000
Model input / output rate	P_in / P_out USD per million tokens, from the pricing page

Session runtime

2,000 sessions × 0.2 h = 400 session-hours × $0.08 = $32.00

Tokens

• Input: 2,000 × 150,000 = 300M tokens → 300 × P_in
• Output: 2,000 × 10,000 = 20M tokens → 20 × P_out

Monthly total = $32 + 300 × P_in + 20 × P_out (+ tool charges)

Two things fall out of the formula:

1. On mid-size and large models, tokens dominate. The runtime fee here is $0.016 per session. On Claude Sonnet 5 ($2 input / $10 output per million tokens) the same session’s tokens cost about $0.40. On Claude Haiku 5.5 ($0.10 / $0.50 for prompts up to 100k tokens) they cost about $0.02, so with a small model the runtime fee is about as large as the token bill. Prompt caching and context size are the biggest levers on larger models.
2. For long-running or always-on agents, runtime becomes a real line item. Ten agents running around the clock are ~$584/month in session runtime only, excluding model token costs.

How the bill changes if you self-host

An open, self-hosted runtime does not make model tokens cheaper. If you run Claude on OpenMA, you pay Anthropic’s token rates directly with your own key. What changes is the shape of everything else:

Cost line	Claude Managed Agents	OpenMA self-hosted	OpenMA hosted
Model tokens	Anthropic rates, Claude only	Your provider’s rates (BYOK), any supported model	Your provider’s rates (BYOK)
Runtime	$0.08 per running session-hour	Your servers, database, and sandbox provider	Wallet-based cloud sandbox compute; see pricing
Platform fee	Included in session runtime	None	None. You pay for sandbox compute (pricing)
Operations	Anthropic	Your team	OpenMA

Self-hosting tends to save money when you move cheaper workloads to other models through the same runtime, reuse infrastructure you already pay for, or keep long-running sessions on compute you have priced yourself. If your team would spend more on operating the platform than the runtime fee costs, CMA is the cheaper option.

For the broader view of options, see Claude Managed Agents alternatives and the shortlist of open-source managed agents.