Claude Managed Agents has two official billing dimensions to track: model tokens and time while a session is actually running. Separately priced server-side tools can add charges on top. As of August 31, 2026, Anthropic lists session runtime at $0.08 per running session-hour. Token prices depend on the selected Claude model, and tools such as web search can add a per-use charge in addition to the tokens they introduce.
Do not combine this bill with Claude Code subscription pricing. Claude Managed Agents is an API platform for running agents on managed infrastructure; Claude Code Pro or Max access is a different product decision. If that is the question you are answering, use the Claude Code pricing comparison instead.
Prices and product contracts can change. The figures below were checked against Anthropic's official pricing documentation on August 31, 2026; verify the current Claude pricing page before committing a budget.
The short answer and cost formula
For one Managed Agents session, use this structure:
Keep a fourth line in your internal budget for costs outside the Anthropic bill, such as your own database, object storage, observability service, custom MCP server, or another vendor API. Those are real workload costs, but mixing them into “Claude token cost” makes optimization harder.
These bill components respond to different controls:
The total is not simply “tokens plus one hour” for every session. Runtime is metered to the millisecond, and only one session status accrues that charge.
Core billing dimension 1: model tokens
Anthropic bills all tokens consumed in a Managed Agents session at the selected model's current rates. That includes ordinary input and output, and the prompt-caching multipliers apply in the same way they do to model requests elsewhere on the Claude Platform.
A useful token ledger separates at least:
- uncached input tokens;
- cache-write tokens, including the cache lifetime used;
- cache-read tokens;
- output tokens;
- tool definitions, tool calls, and tool results that become context.
That separation matters because “input tokens” do not all have one effective price when caching is active. Repeated instructions or a stable repository context may become cheaper to read from cache, while frequently changing context can keep producing cache writes.
Do not assume the Batch API discount applies. Anthropic's Managed Agents pricing section says it does not: sessions are stateful and interactive, and there is no Managed Agents batch mode. If a workload is truly asynchronous, independent, and latency-insensitive, compare it with the Messages Batch API as a different architecture rather than adding a nonexistent discount to a Managed Agents forecast.
Other model modifiers can still matter. The official page says fast-mode premium pricing applies when the agent's model speed is set to fast, and US-only inference applies the documented data-residency multiplier. These settings should be visible in the agent version or deployment configuration, not left as an unexplained difference in the bill.
Core billing dimension 2: running session time
Anthropic currently lists Managed Agents runtime at $0.08 per session-hour, measured to the millisecond. Runtime accrues only while the session status is running.
The following states do not accrue Managed Agents runtime according to the official contract:
idle, including time waiting for the next user message or a tool confirmation;rescheduling;terminated.
This has two practical consequences.
First, a session that exists for an hour is not automatically billed for an hour of runtime. You need the duration of its running intervals, not wall-clock time from creation to termination.
Second, do not add ordinary Code Execution container-hour pricing on top. For Managed Agents, session runtime replaces that container-hour model. Anthropic explicitly says the container is not separately billed in addition to session runtime.
That does not make all infrastructure free. A self-hosted environment, external tool server, data store, network service, or third-party sandbox can still have its own cost. The point is narrower: do not double-count Anthropic's standard Code Execution container rate inside a Claude Managed Agents session.
Additional charges: tools
Tool use affects the bill in two ways. Every tool definition, invocation, result, screenshot, or fetched text that enters model context can add tokens. Some server-side tools also have a separate usage charge.
The clearest current contrast is between web search and web fetch:
- Anthropic lists web search at $10 per 1,000 searches, plus the standard token cost of search-generated content. One search is counted per search execution, not per result returned.
- Web fetch has no additional per-call charge, but fetched content that enters the conversation is still billed as tokens. A large page is therefore not costless.
For custom tools, separate three possible expenses:
- the tokens used to describe the tool and return its result to Claude;
- any Anthropic server-tool fee, where the official price table defines one;
- charges from the external system the tool calls.
This distinction prevents a common diagnosis error. If tool-call count stays flat but token cost rises, the issue may be larger results rather than a higher per-call fee. If token volume stays flat while cost rises, check the selected model, speed, residency, or separately metered tool first.
Build a budget before production
A reliable forecast starts with your own representative tasks, not a universal “cost per agent” benchmark. Task length, tool result size, model, retry behavior, and cache reuse can change the answer substantially.
Run a controlled sample through the intended agent version and record, per completed task:
- model and configuration version;
- uncached input, cache-write, cache-read, and output tokens;
- total time in
runningstatus; - each server-tool request count;
- external tool or infrastructure costs;
- task outcome, including failure, cancellation, or human escalation;
- retries and duplicate work.
Then calculate distributions rather than one average. The median shows the normal path; a higher percentile reveals long sessions, unusually large context, or repeated tool loops that drive budget risk. Keep successful and failed tasks separate so a cheap failure does not make the system look efficient.
For a monthly forecast, multiply each task class by its expected volume, then add a margin based on observed variance. Do not invent a blanket percentage before you have the distribution. A new workflow with little evidence should keep a lower spending cap and a smaller rollout cohort until its tail behavior is visible.
How to lower cost without breaking the agent
Optimize the layer that is actually growing.
If token cost dominates
- Use the least expensive model that still meets the task's acceptance criteria.
- Keep stable instructions and reusable context cache-friendly.
- Limit fetched files and tool results to what the next decision needs.
- Summarize completed phases instead of replaying their entire working history.
- Compare quality and total retries, not just the rate per million tokens.
If runtime dominates
- Find long
runningintervals with no useful progress. - Add explicit tool timeouts and terminate sessions that cannot recover.
- Move human waiting into an idle or approval state supported by the workflow instead of polling inside the agent loop.
- Reduce repeated setup work by versioning a stable environment where that does not create stale-state risk.
If tool cost dominates
- Cap search depth, fetched content, and result size.
- Cache safe, non-volatile external data at the application layer.
- Make retry policy aware of whether the previous call was actually billed or completed.
- Require an expected information gain before repeating an expensive search.
Cost reduction is only successful if task completion and review quality remain acceptable. A cheaper model that causes more retries, or a smaller context that hides a critical constraint, can raise total cost even when the unit price falls.
Managed Agents versus Agent.Space billing
Claude Managed Agents and Agent.Space should not be compared as if they were two identical runtime SKUs.
Managed Agents provides an Anthropic-operated agent loop, agent versions, sessions, sandbox execution, and session events. Its bill includes the token, runtime, and applicable tool layers described above.
Agent.Space has a different boundary. Share supplies fixed OpenAI-model capacity; Flex is a measured balance for other available models; the selected model determines the funding source, and the two do not silently replace each other. The Share versus Flex guide explains that product contract.
Through the Agent.Space Developer API, supported clients can discover available models and use a compatible protocol. That access layer does not claim to reproduce Anthropic's Managed Agents session API, runtime, tools, or pricing.
Ask one question before comparing prices: Do you need Anthropic to operate the agent runtime, or do you already have a coding harness and mainly need model access and a workspace?
- If you need the Managed Agents sessions API and its managed execution boundary, budget the two official dimensions and any applicable server-tool charges directly.
- If you already intend to work in Claude Code or another supported harness, compare the model-access and workspace path separately.
For the second case, review Agent.Space plans using the live model list and current terms. It is an alternative route for a different job, not a promise of feature or billing equivalence with Claude Managed Agents.
The practical takeaway
Keep three ledgers: tokens, running runtime, and billable tools. Add external infrastructure on a separate line, and never add Claude Code subscription fees or ordinary Code Execution container-hour charges merely because the words “Claude” and “agent” appear in the same architecture. Once each cost has its own meter, the correct optimization—and the correct product comparison—becomes much easier to see.