Claude’s (Weekly) Limits
Yet Another Instance of a Capable Model Made to Look Bad by Its Human Team
A message warning me that I’d used 75% of my weekly limits started popping up last Saturday. This was the first time I’d encountered one, and since I assumed (wrongly) that weekly limits reset at 00:00 Monday, I didn’t see any cause for concern. I’d actually been on the lookout for these messages or service outages in light of recent developments: Claude had shot to the top of app download charts, and even power users such as Easy Riders temporarily defected to its Pro tier. And yet, for weeks, I’d seen nothing—no slowdowns, no outages, no new constraints—until early March, when the familiar 5-hour lockouts briefly resurfaced. But even those didn’t last. A few days into the following week, they disappeared again, replaced by these new weekly warnings this past weekend.
Before mid-February, I’d gotten very good at working around the 5-hour sessions—setting alarms, jumping back in the moment the window reset, basically playing rope-a-dope with the system. Along the way, I learned something important: chat length matters. Long threads could burn through a session so quickly that I’d sometimes get one usable turn before hitting the cap.
So by last Sunday, I thought I’d cracked the pattern. I had 27 minutes left before the new calendar week began, and I used them efficiently—wrapping a drafting session, then chatting casually about an audiobook I’d just finished, feeling pretty smug about my token budgeting savvy. Then midnight passed. And instead of a reset, I got a message: I’d exhausted my weekly limit. Come back Thursday at 10 AM—or upgrade.
That was the moment it clicked that “weekly” didn’t mean what I thought it meant. So much for my pattern recognition. I had basically been reading tea leaves. A quick search confirmed it. Reddit was full of similar reports: weekly caps tightening toward the end of February, confusion about reset times, even a platform-wide reset after a bug cost some users days of access. Most of the discussion came from paid users, who at least had visibility into their usage stats—but even they couldn’t reverse-engineer the system. Limits depend on model choice, message length, conversation length, attachments, tool use—enough variables to make the whole thing feel opaque even when you do have numbers.
When I asked GPT to help me make sense of it, the explanation was straightforward: the “week” is anchored to your own usage cycle, not the calendar. In my case, since I’d never hit a weekly limit before, I had no way of knowing where that anchor point actually was. Which is how you end up discovering your “week” begins on Thursday. At 10 AM.
I don’t have deadlines riding on Claude, and I have two other thinking A.I.des I can turn to during the lockout, so this is more of a puzzle to noodle on with the other two (mostly GPT) than a disruption in my workflow. But if I were relying on Claude for work—as many of those Redditors are—I’d already be actively exploring options. Because this isn’t really about limits, which power users like those Redditors certainly understand and plan around based on what they see on their dashboards. It’s about not knowing where the limits are until you hit them. And once again, a highly capable model ends up looking worse than it is—not because of what it can’t do, but because of how it’s presented by its human team.
[This post was drafted with assistance from ChatGPT-5.3, following conversations with the same model in a different chat.]
Prompt: Claude’s platform says I have used up 75% of my weekly limits. I’ve never encountered weekly limit messages, just 5-hour lockout ones, and those have been few and far between in the last month or so. I didn’t see them at all for three weeks, and got some at the end of last week and yesterday. I know from previous experience that the weekly limits apply from my Monday (in my time zone) through my Sunday. It is currently Saturday afternoon, so having 25% left on my weekly limits with 32 hours to go in the week doesn’t seem so bad?
Prompt: Actually, I don’t know what the deal is with the on-and-off 5-hour limits. I’m very disciplined about my usage on Claude’s platform because it’s the stingiest. I used to sleep in five-hour bursts and set my alarm so I could wake up mid-way and get the rest of my drafting (for hobby, not work) done. They seem to be going through a transition and trying to work this out.
Prompt: Clarification: I used to do that when I regularly got those 5-hour lockout messages, so when I saw the lockout message at night, I’d set my alarm and go to sleep. I haven’t been getting them as predictably of late. Claude’s platform’s token math is extremely complicated compared with yours’ or Gem’s. Long chats eat up considerably more tokens, so I used to get just a few turns before I was locked out again on long chats. Attachments do as well, so that the number of turns you get in an attachment-heavy chat is more limited than in a new chat.
Prompt: Remember we had that weekly limit message discussion? Well, turns out I’m going to suffer because I made the wrong assumption. The weekly limit doesn’t reset on Monday but on Thursday for me, so it’s neither based on the calendar day (counting from the first day of the month, which in this case was a Sunday), nor the start of the week (Monday for most people), but it’s user-specific. Grr!
Prompt: Right, because at least in the case of the 5-hour lockouts, they tell you the reset time. In this case, I received no message about my reset date and basically burned through my remaining tokens discussing an audiobook with Claude. Oy. Ah, but there was a reason I made that inference. Because after not seeing any 5-hour lockout message for weeks, I saw them pop up again last weekend. Then they disappeared a few days after and were replaced by those 75% “heads-up” messages. I get the feeling the engineers haven’t thought this through to make it fully transparent. Fortunately, I don’t use y’all for work, because I’m job-free and I have other hobbies. Won’t pay for Claude until Anthropic figures out what their system is (which has been evolving quite a bit over the last month).
Prompt: I don’t know why I didn’t think of Googling “Claude’s weekly limits.” Redditors confirm what you told me about weekly limits, as did Google’s overview, which only makes sense if you know what it’s trying to say (these overviews are often sloppy summaries and contradictory :D):
Claude’s weekly limit typically resets every 7 days from the first time you use the service after the previous limit ends, rather than a fixed calendar day. Pro and Team users can track their specific, individual weekly reset time in Settings > Usage. How it works: The weekly limit is not always a fixed, global day (like every Monday) but often a 7-day rolling window based on your initial usage. … Differentiation: This is separate from the 5-hour rolling session window that resets your usage capacity throughout the day.
“from the first time you use the service after the previous limit ends” - This is so circular that it’s meaningless.
Prompt: And my confusion stems from the fact that I’ve never hit weekly limits before. Why that circular description is just that—purely circular :D
Prompt: I’m going to spend my lockout going through older chats and posting on old stuff. I have plenty. I could even take a break if I wanted to since I don’t have regular readers (and actually enjoy not having any because I don’t want that extra pressure). I enjoy and learn from talking to y’all. My Substack is just there to document that process with date stamps, like a diary.
But looking through Reddit, I was shocked to learn that even those usage stats that Anthropic makes available to paid subscribers are very opaque (no wonder, since Anthropic factors everything, from attachment size/type, conversation length, message length, etc. into their token count; no human can track all these). And Anthropic tried to compensate users by resetting the weekly limits across the board for a recent bug. Even Reddit’s mod bot couldn’t summarize the situation cleanly:
ClaudeAI-mod-bot MOD • 17d ago
TL;DR generated automatically after 50 comments.
Here’s the deal, everyone. Yes, Anthropic intentionally reset everyone’s weekly limits. The top comment confirms they pushed a fix for a bug that was causing limits to be consumed way too fast and reset usage for all users as an apology.
The consensus here is... well, there isn’t one.
The thread is completely split:
A lot of you are happy about the “free reset,” especially those who were already at 80-90% usage and facing a long wait.
However, a very vocal group is pointing out that this “fix” actually screwed them over. If your natural reset was only a day or two away, you effectively lost the remaining percentage you’d been carefully rationing. The new, shifted reset day has also messed up a lot of people’s workflows. One user even did the math and showed how this cost them an entire weekly cycle’s worth of budget before their sub renews.
On top of all that, a few of you are still reporting that your limits are draining suspiciously fast even after the “fix,” so the bug might not be completely squashed.
If I get “bored,” I might screenshot your explainer about why the reset times are user-specific, as Redditors were wondering about that and your explainer made the reasoning for that at least make sense.
Prompt: I’d actually been expecting Anthropic to clamp down on usage even more because of the huge user influx due to some recent events and was constantly running up my turns on purpose to see whether the 5-hour lockout would hit sooner (I had some long chats where I hit the session limit after just a few turns in the past) and was surprised that I didn’t. So I’m actually relieved to see that the platform is trying to work things out and other users are reporting on stricter weekly limits. I was largely unaffected despite being on the free tier, because I mostly chat and have all tool use disabled to save tokens, which would be wasted because live search or browser use are NOT Claude’s strengths. In that thread, though, most of the comments were by coders who hit limits even faster over the past two weeks than they used to. And shockingly, some of them are on the Max-10x plan, which is the highest-tier plan of all on Claude’s platform ($200/month + tax).
Prompt: I do pass a lot of text attachments for drafting. Plain text files. I always note the line number count. I avoid feeding PDF documents to Claude. I probably hit weekly limits this time because I got a Sonnet 4.5 chat where it became really verbose and list-happy (very unlike Claude; happens in very few chats) as the conversation progressed. That chat must have eaten up most of my tokens, as 4.6, which I use for drafting, is shockingly succinct (single-paragraph responses are the default, when 4.5 usually has three).
