Claude Code now uses Haiku 5.5 for background subagent status lines, cutting their usage-limit cost; still no off switch

Anthropic has changed how Claude Code writes the status line for background subagents, and it says the change makes those subagents noticeably cheaper on a subscription. The status text is now written by Haiku 5.5 from short notes, instead of by re-reading the subagent's whole conversation every 30 seconds. According to the ClaudeDevs account, which announced it on November 4, each update now costs about 99% less, and sessions with background subagents used about 5% less of the usage limit, "so the same limit gets you about 6% more prompts."

This is the cost we wrote about on October 29: the agent_summary request behind the three-to-five-word status in the agents panel. Users who captured raw requests found it forked the subagent's full context, sent it to the subagent's own model, often Opus, and asked for a few words. In subagent-heavy sessions they measured it at roughly 16 to 26% of tokens or estimated spend. The cost grew with context size and the number of running agents, not with useful work.

What changed, and what didn't

Per the follow-up post, only the status line moves to Haiku: "The subagent itself keeps its own model." The change is live "for most Pro, Max and Ultra users and for everyone on Team and Enterprise," with Claude Code v2.1.316 or later.

Two details are worth reading closely.

First, the change isn't in the changelog. The release notes for 2.1.316, 2.1.317 and 2.1.318 don't mention Haiku writing summaries; the only summary-related entry, in 2.1.318, is about letting a summary on Bedrock, Vertex, Foundry or with telemetry off wait up to five minutes for a cold prompt cache. Together with "most" users and a minimum version that is already four days old, that points to a server-side switch rather than a new client build. That is our inference; Anthropic hasn't said how it's rolled out.

Second, the 5% figure is Anthropic's average, and it sits between the two camps from last week's debate. It is well above the "~2% typical" number that came up on Hacker News, and well below the 16 to 26% that users measured in heavy sessions. Both can be true: an average over every session with a background subagent includes many short ones, while the people who filed measurements were running dozens of long-lived Opus agents. If you were in the second group, the saving should be larger than 5%. Nobody has published new measurements yet.

Still no off switch

The GitHub issue that asked for a supported way to turn the summaries off is still open, with 12 comments and 120 upvotes since August 15, and as of this writing no reply from Anthropic staff. Making each call cheap is a different fix from letting users skip it. Since our first story, the thread has also become a small reverse-engineering project, and its findings still matter for anyone who wants zero summary calls:

• The undocumented CLAUDE_CODE_AGENT_PROGRESS_SUMMARIES variable can only turn summaries on. Setting it to 0 counts as on, user chrisvaillancourt found in the 2.1.312 bundle.
• CLAUDE_CODE_FORK_SUBAGENT=0 stops them only in an interactive session where nothing else has enabled them. User canneyi traced this through the 2.1.309 to 2.1.312 builds and showed that a settings env block can silently override the shell value, which explains why some people saw it work and others didn't. Fork subagents and skills with context: fork get summaries regardless.
• The timer skips a tick when a subagent's transcript hasn't changed, so the cost follows how often the agent acts, not wall-clock time, GabriellCVig showed from --debug-file logs. The same logs showed the calls weren't pure cache reads: 2.8k to 13k fresh input tokens and 195 to 526 output tokens each on a ~32k context, far more than the ~65 output tokens in the original report.
• In users' tests, claude -p runs produced no summaries, and neither did ordinary background agents in the Claude desktop app's Code tab (forks there still get them). The overhead is mostly an interactive-CLI cost.

The last point is the awkward one. The SDK already treats progress summaries as opt-in. The interactive terminal CLI, where most subscription usage happens, still has no setting for it. Moving the job to Haiku answers the cost complaint, at least on the plans where it is live. It doesn't answer the request in the issue's title.