Claude Usage Increasing While Idle: Separate Meter Changes from Request Execution
If Claude usage increases while you are idle, a rising meter alone cannot tell you which request caused it. Separate three questions: did the counter move, did work execute, and did the client display the result? A limit banner can coexist with reported execution elsewhere, while an idle-usage complaint needs evidence covering the same device and time interval.
Claude’s official usage explanation says its product surfaces share usage limits. A local transcript therefore needs to be read alongside account-level information, rather than treated as the account’s entire history.
On this page
An idle report can contain several different evidence windows
A report filed October 23, 2026 described Claude Code 2.1.309 on Windows 10, Max, with open CLI sessions.
For October 22, it gives a reset at 15:00 and says the process-creation audit was checked from 15:32. Times are the reporter’s local times.
Our reading: the process audit begins after the reported reset. That specific audit cannot establish an absence of new processes at the reset itself. This limits the process-launch claim; it does not erase the separate transcript observations.
For October 23, the day table attributes four weekly points to filtered zero-call intervals. A longer described interval goes from 35% to 42% weekly, with zero calls in most intervals. Its filter excludes intervals containing any local call.
Our reading: subtracting the longer interval’s endpoints gives 42 − 35 = 7 points, but not seven proven zero-call points. The day table and the longer interval use different inclusion rules. Their difference cannot be labelled unexplained idle consumption without the omitted interval records.
The report also describes a nine-minute per-process packet capture and, separately, an 18:00–18:45 zero-call stretch.
Our reading: the short packet capture does not, by its duration alone, witness that entire longer stretch. Its placement must be established before treating it as continuous network evidence for the zero-call interval. Evidence from a transcript and evidence from a packet capture are not interchangeable coverage.
An October 25 update to the Windows report places local authentication failures from 23:16:34 until a new login at 23:47. Its displayed usage comparison is labelled 23:16–23:48. The update specifies UTC+9 for these times.
Our reading: that meter interval extends beyond both ends of the stated authentication-failure interval. It cannot assign the entire counter rise to a continuously logged-out local client. A rise measured wholly inside the failure interval would be a different observation; that paired reading is not supplied here.
“All Claude Code” does not identify one computer
An October 19, 2026 Max 20x report includes these transcript totals for two corresponding windows. The prior window’s weekly meter was not sampled.
| Window, UTC | Calls | Cache write | Cache read | Weekly meter |
|---|---|---|---|---|
| October 11 05:00–October 12 03:16 | 1,103 | 5.36M | 297.6M | Not sampled |
| October 18 05:00–October 19 03:16 | 902 | 7.47M | 299.4M | 0% → 54% |
Our reading: the matching transcript windows do not supply matching meter observations. The missing earlier reading prevents a directly observed, same-duration allowance-consumption comparison. An estimate derived from a longer prior window is a different kind of comparison.
The same account also had a Remote Control session on a second computer. Its usage breakdown assigned 100% to Claude Code, and a remote result included usage_limit_reached with HTTP 429.
Our reading: here, one product bucket contains activity from more than one computer. A 100% Claude Code attribution cannot identify the originating computer. It cannot replace the separate device evidence used in the Windows idle report.
A limit message can be a different problem from refused execution
A June 1, 2026 mobile-versus-desktop report says Extra Usage was active and desktop work continued. Prompts sent from the phone were reportedly received and acted on in the concurrent desktop session, while the phone displayed the old limit message.
Our reading: this mobile account describes observed work despite the banner; the October remote account supplies an error result. Those are different execution outcomes. Combining them as one demonstrated quota failure would discard the very evidence that separates them. Check the existing task’s result before treating the mobile message as proof that the request never ran.
The mobile report is closed as not_planned and labelled invalid in its primary tracker. That status does not itself document a tested fix for the reported client behavior.
What to check before another retry
- Use the named session or weekly counter and its actual observation interval.
- Compare that interval with the coverage of your device evidence; include other active computers and sessions.
- For a cross-device display mismatch, inspect the existing task for execution or a result before resending it.
- Give support the mismatch you can establish: counter movement without matching local activity, or a displayed failure despite observed work. Keep an unverified cause unresolved.
These examples do not establish whether your account is being charged incorrectly, whether a current client still has the reported behavior, or whether a refund is due. They explain how to avoid overstating the available evidence. For a different issue—matched work consuming different session shares through different entrypoints—see the Agent SDK usage comparison.