Claude Code’s background subagents now use Haiku 5.5 to write status updates from short notes instead of rereading the full subagent conversation every 30 seconds, according to Anthropic’s developer account, ClaudeDevs. ClaudeDevs estimates that this makes each update cost about 99% less and that sessions with background subagents use about 5% less of the usage limit. Those estimates describe different things: the cost of one status update and usage across a session.
What the status-line change does
ClaudeDevs said the change was live for most Pro, Max, and Ultra users and all Team and Enterprise users, on Claude Code v2.1.316 or later. The subagent keeps its assigned model; Haiku 5.5 handles the status line, not the delegated work.
The reported savings apply to status updates in sessions with background subagents. They do not give subagents a separate usage allowance.
How to read the savings estimates
The approximately 99% reduction is ClaudeDevs’ estimate for each status-line update. Its separate estimate of about 5% less usage applies to sessions with background subagents. The account also described the same limit as allowing about 6% more prompts.
Those are rounded estimates, not a guarantee that every subagent workflow uses less. As a simple calculation, exactly 5% less usage per prompt would allow about 5.26% more prompts from the same quota, assuming the work and other factors stay the same: 1 ÷ 0.95 − 1 ≈ 0.0526. That calculation does not make the company’s approximate figures exact.
The status-line change is also separate from the earlier reported adjustment to Claude Code’s weekly allowance, covered in NeoTeo’s earlier Claude Code coverage. The November announcement concerns background status-update costs, not a broader change to account quotas.
Subagents use the main conversation’s allowance
Anthropic’s Claude Code subagent documentation says subagent requests count toward the same usage limits as the main conversation. Running agents in parallel therefore does not create a separate pool of allowance; their requests still count against the shared limit.
The status-line optimization addresses one recurring cost within that usage. It does not mean every task delegated to subagents reduces total usage: the subagents’ own requests remain part of the account’s consumption.
Controls for subagent runs
Claude Code provides controls over a subagent’s turns and over how many agents can run or spawn other agents. The Agent SDK also offers a query-level spend cap that includes subagent requests.
| Control | What it limits | Scope and default |
maxTurns | Agentic turns for one subagent | Per subagent; reaching the cap returns partial output that can be resumed. |
CLAUDE_CODE_MAX_CONCURRENT_SUBAGENTS | Subagents running at the same time | Agent-tool subagents; default maximum of 20. |
CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH | How deeply subagents can spawn other subagents | Default depth of three layers below the main agent. |
maxBudgetUsd / max_budget_usd | Spend for a query, including subagent requests | Agent SDK query-level cap, not a per-subagent Claude Code CLI budget. |
Anthropic’s documentation sets no cap on the total number of subagents spawned across a session. The default of 20 limits concurrent Agent-tool subagents, not total session spawns. maxTurns limits turns, not tokens or dollars; the SDK’s spend cap applies to a query rather than to an individual CLI subagent.
