Status lines now written from short notes
Anthropic said on Nov. 4 (Nov. 5 Japan time) that it had reduced how much of the usage limit is consumed by helper AIs run in the background (background subagents) in its coding tool "Claude Code." The status line that shows what is happening during a run is now written by the small model "Haiku 5.5" from short notes. According to Anthropic, the cost of each update falls by about 99%. For work that uses these helper AIs, consumption of the usage limit falls by about 5%, allowing about 6% more prompts within the same limit, it said.
Background subagents in Claude Code now use less of your usage limit. A running subagent updates its status line every 30 seconds, and each update used to re-read its whole conversation. Now Haiku 5.5 writes it from short notes instead, so each update costs about 99% less.
Subagents are helper AIs that split up and carry out parts of a task, running separately from the conversation with the user. The status line shown beside them is a short phrase of three to five words, such as "running tests." According to the developer blog dsebastien.net, the line had previously been generated by the subagent's own model, which was usually the higher-end "Opus."
According to AlphaSignal, under the new method the subagent updates short notes as it works, and Haiku 5.5 reads those notes instead of the whole conversation. Even as the conversation grows longer, the cost of updates is unlikely to rise. The model running the delegated work itself does not change.
About 99% less per update, but about 5% savings overall
The figure of about 99% applies to each status update, and savings across the whole task are only about 5%, according to Anthropic. That is because status updates are only part of a task's overall consumption. Anthropic has not explained how it calculated the figures of about 5% and about 6%. According to AlphaSignal, the effect varies with the length of the task, the length of the conversation and the number running at once, and there is no effect on work that does not use background subagents.
No third-party measurement of the effect after the change had been found as of Nov. 5.
Developer measured 16.4% of costs; Anthropic says about 2% on average
Consumption by status lines became a topic among developers in late October. A developer calling themselves Flamerix AI published measurements showing that status updates accounted for 16.4% of a week's costs and 19.7% of requests. Of these, 96% were processed by Opus, and input tokens (units of text) were about 9,000 times the output. The costs are Claude Code's own estimates and differ from actual billed amounts. The post drew 589 points on Hacker News.
edwinarbus, who identified himself as a Claude Code developer, acknowledged on Hacker News that the explanation was broadly correct but countered that the share was lower. He said it was about 2% of costs over the most recent week (converted at API list prices) and just under 4% of requests, and about 6% of costs when limited to the terminal. He regarded the 16.4% as an example from a team that mainly uses background agents and said it was not typical. He disclosed at the time that a cheaper method that sends short notes to a small model was in development.
The roughly 6% for the terminal is a share of costs and is a different figure from the "about 6% more prompts" in this change.
Rolled out to most subscribers; no official setting to turn it off confirmed
According to Anthropic, the new method has been rolled out to most Pro, Max and Ultra users and to all Team and Enterprise users. According to Anthropic, it applies to v2.1.316 and later, and no settings change is needed. The 2.1.316 entry in the official changelog does not mention this change.
ORIENS could not confirm which users are excluded. Anthropic's announcement also does not mention use through other companies' clouds, such as Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry. The changelog for 2.1.318 on Nov. 4 addresses, in a separate entry, status lines for use through those routes and when usage reporting (telemetry) is turned off.
A request for an official setting to turn off the status line itself was filed on GitHub on Aug. 15. It reported a case in which 27 subagents were run over four hours and status lines accounted for 18.0% of weighted tokens. The request remains unresolved, and there was no response from Anthropic in comments through Nov. 3. The developer calling themselves Flamerix AI reported that 2.1.318 has an environment variable to turn off status lines, but it is not in the changelog.
Responses to complaints about usage limits continue
Claude Code has faced continuing complaints that usage limits are depleted faster than expected. Boris Cherny, who leads Claude Code, reported on Oct. 28 (Oct. 29 Japan time) that the team was fixing multiple small bugs that accelerated consumption and said no single cause had been found. Anthropic has not said whether this change is one of those fixes.
In October, Anthropic added features to reduce consumption, such as automatic summarization of long conversations left idle. In 2.1.318 on Nov. 4, it moved commands run by subagents to the background after three minutes, so that the conversation-reuse mechanism (prompt cache) does not expire while they run.
OpenAI also introduced savings measures in its coding tool "Codex" on Oct. 16 (Oct. 17 Japan time), such as reusing context among subagents, but did not give figures for the size of the improvement.
- Status lines enabled by default in 2.1.232Third party
- Request on GitHub for a setting to turn them offThird party
- Developer publishes measurement of 16.4% of costsThird party
- Person identifying as a Claude Code developer counters with about 2%, describes cheaper method in developmentInvolved party
- Cherny reports findings on usage limitsOfficial
- New status line method announcedOfficial
