Short version: Claude Managed Agents bill on two meters, not one — tokens at the usual per-model rates, plus session runtime at $0.08 per session-hour. The runtime meter runs only while the session's status is running; idle, rescheduling, and terminated time is free. And the discount playbook you use on the raw Messages API is mostly switched off here: Batch, Fast mode, data residency, and partner-cloud pricing don't apply to Managed Agents. The one lever that still works — and works well — is prompt caching.
Two meters, added together#
Most people cost a Managed Agent the way they cost an API call: tokens times rate. That's half of it. A Managed Agents session is billed on two independent dimensions that get summed:
- Tokens. Every token the session consumes is billed at the standard per-model rate from Anthropic's pricing page, and prompt-caching multipliers apply identically. Web search triggered inside a session adds the standard $10 per 1,000 searches.
- Session runtime. A flat $0.08 per session-hour, metered to the millisecond.
The second meter is the one that surprises people, so it's worth being precise about when it runs.
The runtime meter only runs when the agent does#
Session runtime accrues only while the session's status is running. Time the session spends idle — waiting for your next message, or for a human to confirm a tool call — does not count. Neither does time spent rescheduling or terminated.
That single rule changes how you should architect a long-lived agent. A session that does a burst of work and then waits for input bills for the burst, not the wait. So an agent that parks cheaply between turns — going idle instead of being held hot — costs almost nothing on the runtime axis, even if the conversation spans hours or days. Design for that: return control and let the session go idle rather than spinning while you wait on an external event.
One more thing this replaces: for Managed Agents, session runtime takes the place of code-execution container-hour billing. You are not billed for both. If you've been mentally adding container hours on top, drop that line item.
The four discounts that don't apply#
Here's where cost models built on the raw Messages API quietly break. Four modifiers you lean on elsewhere are switched off for Managed Agents sessions:
- Batch API 50% discount — gone, because sessions are stateful and interactive; there's no batch mode to run them in.
- Fast mode premium — not applicable; inference speed is managed by the runtime, so you neither pay the premium nor choose the mode.
- Data-residency 1.1x multiplier — doesn't attach, because
inference_geois a per-request Messages API field, not a session one. (Managed Agents residency is governed by workspace geo instead — see the inference_geo breakdown.) - Partner cloud pricing — not available; Managed Agents aren't offered on Bedrock or Google Cloud, so their regional pricing never enters the picture.
If you were counting on Batch to halve a bulk agent workload, that plan doesn't survive contact with Managed Agents. Model the cost at standard rates.
What still cuts the bill: caching#
The lever that does work is prompt caching, with the same multipliers as everywhere else — a cache read costs 10% of the standard input price. For an agent that carries a large, stable context (system prompt, tool definitions, retrieved documents) across many turns, that's the difference-maker.
Anthropic's own worked example makes it concrete. A one-hour Opus 5 session using 50k input and 15k output tokens:
| Line item | Calculation | Cost |
|---|---|---|
| Input tokens | 50,000 × $5 / 1M | $0.25 |
| Output tokens | 15,000 × $25 / 1M | $0.375 |
| Session runtime | 1.0 hr × $0.08 | $0.08 |
| Total | $0.705 |
Now serve 40k of that input from cache instead of reprocessing it every turn:
| Line item | Calculation | Cost |
|---|---|---|
| Uncached input | 10,000 × $5 / 1M | $0.05 |
| Cache reads | 40,000 × $5 × 0.1 / 1M | $0.02 |
| Output tokens | 15,000 × $25 / 1M | $0.375 |
| Session runtime | 1.0 hr × $0.08 | $0.08 |
| Total | $0.525 |
That's a ~25% cut, and notice where it comes from: almost entirely the input side collapsing from $0.25 to $0.07. The runtime meter ($0.08) barely moves the total — which is the real lesson. On a Managed Agent, your bill is dominated by tokens, not wall-clock, and the fastest way to shrink tokens is to cache the parts of the context that don't change.
The founder read#
Budget a Managed Agent as tokens + $0.08/session-hour, and cost the tokens at standard rates — Batch, Fast mode, and data-residency discounts aren't in play. Then spend your optimization time in the two places that pay: prompt caching (cache the stable context; it's the only big discount left) and session lifecycle (let sessions go idle so the runtime meter stops). Get those two right and the second meter stays a rounding error — exactly where you want it. For the wider question of which model to put behind each step in the first place, the tiered-router math still governs the token half of this bill.
