Weekly limit went to 35% in one night without ever hitting the 5-hour limit. I pulled the token counts from my transcripts and that night wasn't even my heaviest day.

Max plan, Claude Code CLI 2.1.292. For weeks the rule of thumb on my account was: hit the 5-hour limit three times and the weekly "all models" bar sits at roughly 10 to 15 percent. Last night, from 21:30 to 10:00 local, I ran one main session plus subagents on Opus 5.5 (up to three in parallel, effort xhigh). The 5-hour limit was never reached. No limit message, nothing paused. This morning the weekly bar stood at 35 percent. That cannot work with the old ratio, so I measured what actually ran. Every assistant message in the local transcripts carries its usage fields. Summed over all sessions of the last three weeks: Night of Oct 9/10 (12.6 h): 3 396 API calls, 116 M cache creation, 1 501 M cache reads, 1.63 M output Oct 7 (24 h, all Opus 5.5): 5 469 calls, 120 M cache creation, 2 534 M cache reads, 3.05 M output Oct 5 (24 h): 7 076 calls, 76 M cache creation, 2 754 M cache reads, 4.63 M output Oct 2 (24 h): 11 169 calls, 52 M cache creation, 4 164 M cache reads, 5.17 M output The heaviest rolling 5-hour windows of the three weeks are on Oct 2, Oct 5 and Sep 23, not in that night. On none of those days did the weekly bar jump like this. So either the counting changed between Oct 7 and Oct 9, or something is off in how subagent usage lands on the weekly bar versus the 5-hour bar. The docs say subagent requests count toward the same limits. They also say cached tokens count less. Neither explains a 35 percent week from a night that was lighter than three earlierHas anyone else seen the weekly bar move much faster since this week? I have sent the numbers to support and will post what they say.