Managed Agents bills $0.08 per session-hour; Sume's cap is separate

Claude Managed Agents charges tokens plus $0.08 per running session-hour. How that meter relates to Sume's max_spend_usd and spend cap.

5 min readSume
All posts

Claude Managed Agents bills two things: tokens at the model rates, and session runtime at $0.08 per session-hour, counted only while the session status is running. If such an agent calls Sume's hosted MCP server, you have three meters running: Anthropic tokens, Anthropic runtime, and Sume's generation spend. Only the last is capped by Sume's max_spend_usd or, for Agent Completions, generation_spend_cap_usd.

Anthropic's meter

The pricing page says runtime is measured to the millisecond and accrues only while the session is running. Time spent idle, rescheduling or terminated does not count. Tokens are billed at the model prices, prompt-caching multipliers apply, and the Batch API discount does not, because sessions are stateful and interactive. Web search inside a session costs $10 per 1,000 searches.

Claude Managed Agents meters, Anthropic pricing page, read 2026-10-08
MeterRateNotes
Session runtime$0.08 per session-hourOnly while status is running
TokensModel ratesCache multipliers apply
Batch discountNot availableSessions are interactive
Web search$10 per 1,000 searchesPlus token costs

Where the time goes with Sume

A Sume render can take minutes, and the agent waits with jobs_wait, which holds for up to 55 seconds per call on the remote server. While a tool call is in flight the session is working, so the wait is runtime. At $0.08 per hour, ten minutes of waiting is about one cent. The runtime meter is small next to a paid generation, which is the reason the Sume cap deserves the attention.

The Sume cap is a different kind of control. max_spend_usd is optional on hosted MCP and enforced only when you give it, so a managed agent should send it on every paid call. idempotency_key is required on paid and write calls. For unattended runs Sume's Agent Completions make generation_spend_cap_usd mandatory.

What to set

See auto permission mode and paid Sume calls in Managed Agents for the approval side.

  • Set a per-call max_spend_usd from a budget variable, not from model text.
  • Tell the agent to poll in slices and never to resubmit a create after a timeout.
  • Track the two bills separately; a session that is cheap in runtime can still hold an expensive Sume job.
  • Use a read-only OAuth scope for sessions that only need to look at jobs.

A budget sketch with published numbers

Anthropic's worked example prices a one-hour Opus 5 session with 50,000 input and 15,000 output tokens at $0.705 in total, of which $0.08 is runtime. Runtime was about 11 percent of that example. For a cheaper model the share would be higher, and for a session that waits on a long render with few tokens it can dominate the Anthropic side of the bill.

A Sume job is billed on top. Treat that as a fixed line you set with the cap, and treat the Anthropic side as the part that scales with how long the session stays running.

Sources

Related posts

More in Pricing

All Pricing posts

Written by Sume