Claude Code usage limit: background agent statuses now 99% cheaper

On the evening of November 4, the ClaudeDevs account announced on X that background subagents in Claude Code now eat less of your usage limit. Each agent’s status refreshes every 30 seconds, and until now every update re-read the agent’s entire conversation. Now Haiku 5.5 writes it from short notes, so each update costs about 99% less.

What changed in how Claude Code counts the usage limit?

This is about the short status line we see next to every agent running in the background. According to the ClaudeDevs post, the mechanism used to work like this:

1. Every 30 seconds, separately for each running agent, Claude Code generated a new status.
2. Each of those updates re-read the agent’s entire conversation, which keeps growing over time.
3. Now the status is written by Haiku 5.5 from short notes instead of reading the whole transcript.

According to the Claude Code changelog, Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens in the API.

The post spread fast: by the morning of November 5 it had more than 16,000 likes and nearly 1.2 million views. No surprise, since limits are one of the most common topics among people who work with Claude Code every day.

Why does it matter?

Because this exact mechanism was behind the much-discussed measurements from a week ago. I wrote about them in my piece on Claude Code token usage: according to a gist by Flamerix AI, agent_summary summaries accounted for 16.4% of the estimated session cost, more than the main conversation.

Since then, more analyses have landed in issue #86879. On October 31, one user calculated from version 2.1.314 logs that a single summary sent between 2.8k and 13k uncached input tokens, and in their model of a week of multi-agent work, summaries made up about 13–32% of spending. Others posted their own workarounds, from environment variables to scripts loaded at startup.

Interestingly, nobody from Anthropic has commented on the issue itself (yet). As of November 5 it is still open, and the change was announced on X, not in the GitHub thread.

What else was improved for background agents?

The changelog for version 2.1.318 from November 4 lists several changes in the same area:

1. Background agents move a shell command to the background after 3 minutes on all providers, so the prompt cache doesn’t expire while it runs.
2. On Bedrock, Vertex and Foundry (and with telemetry disabled), the agent progress summary can wait up to 5 minutes for the agent’s reply when the cache is cold.
3. Fixed the “Now using usage credits” message that appeared about every 30 seconds while a background agent was running.

The ClaudeDevs post itself gives no version number and doesn’t mention needing to update. The changelog doesn’t describe this change directly either, so the simplest thing is to just run the latest version (let’s check it with claude --version).

What to watch out for?

First, the scale. The 99% applies to a single status update, not the whole bill. How much we actually save depends on how many agents we run at once and how long they work. Sessions without background agents shouldn’t notice the change.

Second, there’s no off switch. The post doesn’t mention a setting to turn statuses off, and users are still asking for one in the issue.

Third, the numbers from the measurements. The earlier 13–32% shares are user estimates (from the cost_usd field and cost models), not invoices, so let’s take them with a grain of salt. The best way is to check it on your own setup with OpenTelemetry telemetry, grouping requests by query_source, as I described in my previous piece. For more on working with agents, see the Claude Code tutorial.

FAQ

Why does my Claude Code usage limit run out so fast?

Every request sends the conversation context to the model, so long sessions and many agents at once burn through the limit quickly. According to user measurements, background agent statuses also generated a large share of the cost, and since November 4 they are about 99% cheaper per update.

Do I need to turn anything on to benefit from the change?

According to the ClaudeDevs post, there’s nothing to configure. It’s still worth running an up-to-date version of Claude Code.

Can I turn off background agent statuses entirely?

There’s no official setting for now. User workarounds are circulating in issue #86879, but they don’t work everywhere and aren’t supported by Anthropic.