[BUG] Usage jumped from 80% to 100% for the week.

I just jumped from ~50-60% (iirc) to 95% in the last hour, on a single opus 4.8 xthink session, with 7% usage in the 5hr window, and only 400.8k tokens in the context window for this session. I've not run a single other session in that time, so something does appear to be wrong...

Maybe they silently tweaked how much the weekly usage is worth and applied it retroactively?

CC 2.1.181, on linux (archlinux x64, 7.0.12-zen1-1-zen), Max x20 plan

The piece that explains a lot of these "weekly jumped but my 5-hour window barely moved" reports: **the weekly pool is charged on total tokens processed, and your input context is re-sent every turn** — so weekly burn tracks roughly Σ(context size) across turns, not just the new text you type. A session sitting at ~400k context (like the comment above) bills a large chunk per turn against the weekly pool even when each turn adds little. Prompt caching discounts the unchanged part, but extended thinking (xthink) and any cache misses add real tokens on top. The 5-hour window tends to track session intensity, so a big-context or heavy-thinking session can move the weekly number fast while the 5-hour % stays low — which is exactly the asymmetry you're both seeing.

Practical mitigations if that's the cause: keep context small (/compact often), avoid long xthink runs on an already-large context, and CLAUDE_CODE_DISABLE_1M_CONTEXT=1 so a session doesn't silently grow into the 1M tier where every turn reprocesses far more.

That said — the original report (80%→100% in ~1 minute, 5-hour window at 23%) is hard to account for by reprocessing alone: a single minute of interactive use shouldn't be able to consume ~20% of a weekly pool unless something batched or a counter glitched. If you can confirm it really was ~1 minute from a not-huge context, that's worth flagging as a possible accounting bug rather than expected reprocessing — ideally with the session's context size at the moment it jumped, since that's the number that determines per-turn weekly cost.

Same here, just one message on 4.8 jump from 60% to 95%.

It was approx ~5 minutes, the maximum context usage I have is 17%/1M. Subagents might have been active, but I don't have anything ridiculous like dynamic workflows etc which would have burned through tokens. But even in that case I would have expected the 5-hour tokens to get expired first, the fact that it didn't points to some sort of counter reset.

I just watched my weekly quota jump from 91% to 100% without me even using any claude agents. Was there some kind of retroactive adjustment to usage? I can't find any info on this anywhere.

@yurukusa

the weekly pool is charged on total tokens processed, and your input context is re-sent every turn

Have we got a source for this?

so weekly burn tracks roughly Σ(context size) across turns, not just the new text you type

The issue is that, as you then noted yourself in the third paragraph, it's unrealistic that it should cause a spike like that, particularly since I have had sessions with larger single context windows burn less weekly tokens. This is very clearly an outlier.

so a session doesn't silently grow into the 1M tier where every turn reprocesses far more

Again, source? I thought the cost difference between 256k and 1m was removed quite a while ago... So that part should be largely irrelevant.

The same bug are 30min ago. Almost instantaneous depletion of ~30% of the remaining weekly limit

Same! I go 30% left on max 20x and suddenly 0%. I can't use it right know. Please fix it asap.

same my 40 percent got done in a blink, please fix asap, we are blocked

I just encountered the same bug: my weekly limit usage jumped from 50% to 80% (at max x20) within a 30-to-90-minute window. I just returned to my PC after being AFK and saw a ~30% limit drop in a single hour. For context, my usual daily usage is only 20-30% across a 10+ hour workday. There were no extra heavy tasks running, and checking 'ccusage' via Claude confirmed that everything is normal with no unusual activity.

Same here and the context window feels much smaller,
in 40 minutes my 5h session is done now. I use pro 1x

same here - not sure if related to 2.1.181 or #69313

Yep mine went from 44% to 88% within a few minutes. This happened about 3 hour ago.

Looks like they reset the limits completely 🤷🏾‍♂️

You're right to push on both — correcting my earlier comment:

1. 256K vs 1M cost: you're correct. Current Opus models (4.6+) serve the full 1M context window at standard pricing, with no long-context premium. My "grows into a pricier 1M tier" point doesn't apply — scratch it.
2. The spike: the part I can actually ground is just the documented API behavior — the API is stateless, so each turn re-sends the full conversation context, and prompt caching offsets most of that (cache reads ≈ 10% of input cost). That makes "long context × many turns" burn faster than the new text you type would suggest. But how the weekly subscription quota meters that isn't something I have a public source for — and as you say, it doesn't explain a jump while the 5-hour window barely moves, or usage climbing with no active session. That looks like a provider-side metering issue, not context size. I shouldn't have implied the mechanism covered the outlier. Keeping this open for Anthropic on that basis makes sense.

Apologies for the half-baked first pass.