Claude Code stopped burning tokens on subagent status updates: the update is 99% cheaper

Imagine hiring an assistant to review code, only for them to reread the entire work chat from the beginning every 30 seconds just to tell you, "still working." That is roughly how background subagents behaved in Claude Code. Anthropic has fixed it, and the news matters to anyone paying for tokens.

What was wrong

Claude Code's background subagents handle long-running tasks: reviewing code, running tests, and compiling technical reports. To show you what was happening, they updated their status every 30 seconds. The problem was how those updates were generated. For every update, the base model reloaded the entire conversation history—thousands of tokens, or hundreds of thousands in long sessions.

Context grows quickly in sessions like these, so a single line such as "tests are running, 40% complete" could be disproportionately expensive. Users online calculated that these technical reports consumed up to 16% of the token budget. Anthropic confirmed that the mechanism did indeed work this way.

How it was fixed

Status updates are now generated differently. The agent keeps short notes on its progress, and Haiku 5.5 uses them to write the status text. It is a lightweight, inexpensive model that does not need the full context. A few lines are enough to summarize "what is happening right now."

According to Anthropic, the figures are as follows:

• the cost of a single status update fell by roughly 99%;
• overall savings when working with background agents are around 5%;
• this is equivalent to roughly 6% more requests under the same limit.

The gap between 99% and 5% is easy to explain. Status updates accounted for only part of the spending; the rest still goes toward the agent's actual work: reading files, making edits, and running commands. The 16% reported by some users was most likely an extreme case involving very long sessions. For typical workloads, 5% looks like a fair estimate.

What to do in practice

First, update Claude Code to v2.1.316 or later. The fix will not activate by itself while you remain on an older version. Second, keep the limitation in mind. The optimization works only in the command-line interface; the desktop app has not received it. If you run long background tasks in the desktop version, you will not see the savings yet.

Third—and most useful for developers and product teams building their own agents—this story offers a good design pattern. Not everything the user sees needs to pass through the most expensive model with the full context. Status messages, labels, brief summaries, and notifications can easily be produced by a small model from compact notes. If your agent periodically "reports on its progress," check how many tokens those reports consume. You will probably find the same kind of overhead.

What remains unclear

The figures should be viewed soberly. They are Anthropic's own estimates, and there are no independent measurements yet. It is also unclear how the quality of the status updates will be affected: shorter notes can mean lost details. That is acceptable for a progress indicator but not for diagnosing a failure, so in problematic cases it is better to inspect the full log rather than a summary.

The conclusion is simple. This is not a headline-grabbing release, but a fix for an inefficiency that users spotted themselves. Findings like this show that the cost of agentic systems is often consumed not by the "smart" work, but by operational overhead. That is what should be optimized first—including in your own products.