Anthropic's Claude Code Slashes Background Agent Status Costs by 99%

Anthropic changed how Claude Code generates status lines for background subagents, reducing the token cost of each refresh by about 99%, according to the company’s announcement. Claude Code v2.1.316 uses compact running notes to produce those updates, eliminating repeated scans of a subagent’s growing conversation.

Developers who run several long-lived background agents should see affected sessions consume about 5% less of their usage allowance. Anthropic estimates that reduction allows roughly 6% more prompts within a limit window. The optimization covers status-summary calls and leaves each delegated task on its configured model.

A one-line status carried full-context cost

Claude Code subagents are delegated workers with their own context, tools, and model. Developers can assign them scoped jobs such as exploring a repository, reviewing a diff, running tests, or preparing a pull request. Background execution lets those jobs continue while the main session handles other work.

While a background subagent runs, Claude Code refreshes a one-line status update every 30 seconds. Each refresh previously sent the worker’s entire conversation to a model for summarization. As the transcript grew, every update required more input tokens, even when the resulting status remained a single line.

Those summary calls counted against the same rolling five-hour and weekly allowances used by prompts, file scans, tool calls, and extended thinking. A limit guide explains how those windows affect Pro and Max sessions, while a separate usage analysis examines the cost of long contexts and subagents.

Running notes replace transcript scans

Version 2.1.316 gives each subagent a compact set of notes that it updates while working. Haiku 5.5 reads those notes to generate the status line. Because the summary input remains bounded as the conversation expands, Anthropic says each refresh costs roughly 99% less.

Bounded input changes the scaling

Full-transcript summarization made status costs rise with conversation length. Incremental notes remove transcript length from that cost curve, leaving the number of active agents and refreshes as the main variables. The approach also avoids rebuilding a summary from the complete conversation every 30 seconds.

Anthropic assigns Haiku 5.5 the narrow job of turning compact notes into a one-line update. Developers can still choose another model for delegated work, including through CLAUDE_CODE_SUBAGENT_MODEL. A reviewer example shows Haiku handling a read-only code-review task. Status optimization is automatic and independent of that setting.

Long, parallel runs save the most

Anthropic’s overall estimate applies to sessions that use background subagents, and actual savings depend on task duration, transcript size, and concurrency. Workflows without background subagents receive no usage reduction from this change.

• Long autonomous tasks: Agents that run for many minutes trigger repeated status refreshes.
• Parallel workflows: Multiple active workers multiply the number of summary calls.
• Transcript-heavy jobs: Repository exploration, review, and test loops previously made each refresh progressively more expensive.
• Rate-limited sessions: Pro and Max users near a rolling limit should gain the clearest practical benefit.

Version 2.1.316 enables the change

The update is available to most Pro, Max, and Ultra users, as well as everyone on Team and Enterprise plans. It requires Claude Code v2.1.316 or newer and introduces no configuration changes.

1. Upgrade Claude Code with the package manager or installer used for the existing installation.
2. Run claude --version in a terminal.
3. Confirm that the reported version is 2.1.316 or later.