Anthropic fixes token overconsumption in Claude Code

Anthropic confirmed and fixed token overconsumption in Claude Code: background subagents were updating the status line every 30 seconds, forcing the base model to re-read the entire conversation each time. Now, the status line is written by the cheaper Haiku 5.5 model based on short notes.

Compiled by Sergey KostenchukPublished 2026-11-06Updated 2026-11-06

2026-11-06 Coding Anthropic

🛠 Anthropic fixes token overconsumption in Claude Code

Background subagents performing code checks or tests were updating the status line every 30 seconds, causing the base model to re-read the entire conversation. Now, the status is written by the cheaper Haiku 5.5 model based on short notes, without the full context. The fix is available in v2.1.316, CLI only.

🌍 Up to 16% of an agent's token budget was spent on a single-line status—operational tasks instead of useful work. Offloading routine tasks to a cheaper model is an orchestration pattern that reduces costs for long sessions and multi-agent pipelines.

👤 Update Claude Code to v2.1.316: with the same limit, this results in approximately 6% more requests. 99% is the savings per status update, not per session; the gain per limit is about 5%.

Source 1: https://x.com/ClaudeDevs/status/2118063143167693011 Source 2: https://daily.dev/posts/claude-code-background-subagents-now-cost-about-99-less-per-status-update-ocseldjtf

Sources