Anthropic has changed how Claude Code behaves when it hits its 5-hour usage limit. According to a September 25 statement from the official ClaudeDevs account, when Claude Code runs into the 5-hour cap mid-task, it now looks for a reasonable stopping point and wraps up there, instead of getting cut off with a file half-edited.
The usage needed for that wrap-up is paid for out of a small fixed allowance carved from the user's weekly quota. The feature is rolling out in stages: Pro users get it once a week, while Max and Team Premium users get it every time they hit a 5-hour session limit. If someone wants to keep working after the wrap-up, they have to switch to extra usage — paid on top, by consumption.
The Cost of Being Cut Off Mid-Task
Developers who've run long tasks through Claude Code know the old behavior well: the moment the limit hit, the session stopped dead, even mid-edit on the second change in a third file. The codebase was left half-modified, tests wouldn't pass, and the next session had to start by reading through the diff just to figure out what was finished and what needed to be rolled back.
For a coding agent, that kind of interruption is far more costly than it is for a chat product. Cut off a chat, and you lose part of an answer. Cut off a coding agent, and you lose the consistency of the workspace. That's exactly what the wrap-up mechanism addresses: it lets the model finish the edit it's in the middle of before stopping, leaving the repository in a state that compiles and can be handed off.
ClaudeDevs' statement didn't spell out exactly what the wrap-up involves — whether it runs a test pass along the way, or leaves behind a note on progress. That will only become clear once the feature is actually out and developers can see it in action.
A Season of Tweaks to the Limits
Line up the timeline, and Claude Code's usage limits have been in near-constant motion this year.
On May 6, Anthropic doubled the 5-hour limit and added a temporary 50% boost to the weekly quota, an increase later extended through September 14. Starting September 14, the weekly quota became permanently 25% higher than the old baseline — which, compared with the temporary boost period, actually cut available weekly usage by roughly a sixth.
On September 25 — the same day the wrap-up feature was announced — the Claude developer blog published a breakdown of what it costs to run Opus 5.5. At list API pricing, Opus 5.5 runs $4 per million input tokens and $20 per million output tokens, both 20% cheaper than Opus 5; cached-read pricing dropped from $0.50 to $0.20, a 60% cut. For the same token volume, the post put a single session at roughly 31% cheaper; Anthropic's own line is that running costs under typical load are 40% lower than Opus 5. But Opus 5.5 always thinks before it answers, so actual token usage varies by task, and the post recommended users test it against their own workloads.
The same post also touched on the usage limits:
"Five-hour limits went up on Pro, Max, Team and seat-based Enterprise plans."
Eligible subscribers can also claim a one-time usage reset from the usage page in settings.
What's Still Unclear
Taken together, Anthropic's moves on usage limits fall into two buckets: pushing unit prices down, and fine-tuning the limit rules themselves. The wrap-up mechanism belongs to the second bucket — it doesn't add to the total usage available, it changes what happens at the exact moment usage runs out.
Three things remain officially unconfirmed. First, how big that "small fixed allowance" actually is, and whether it's measured in tokens or in time. Second, since the allowance is deducted from the weekly quota, whether an already-tight weekly limit will now run out even sooner. Third, what happens if a Pro user hits the limit more than once in a week — the statement only says "once a week," without addressing what comes after the first time.
For developers in China who use Claude Code through a Pro subscription, a once-a-week allowance means timing is something they'll have to manage themselves. Scheduling the changes most likely to get cut off for a moment when quota is plentiful is a safer bet than counting on the wrap-up mechanism as a backstop. Max users get a wrap-up every time they hit the limit — the difference will show up on long-running tasks.
Sources: ClaudeDevs' official account, CocoLoop, and the Claude developer blog post "What a Task Costs on Opus 5.5." Opus 5.5 vs. Opus 5 API pricing and discount figures are verified against the developer blog's own numbers.