Claude Max Weekly Limit: Our Usage Fell by Half in September 2026

The Claude Max weekly limit looks much smaller in our own data since late September 2026. On a Max 5x account that hit the weekly limit every week, output tokens per week fell 63% against the August average, total tokens fell 48% and API-equivalent value fell 62%. This is one account (n = 1), so treat it as a case study. It does not prove that every account changed. The only weekly-limit change we found announced by Anthropic in this period is a 17% reduction on 14 September, which is much smaller than what we measured.

What does our data show?

We use Claude Code on a Claude Max 5x plan, and the account reached the weekly limit in every week below, so weekly usage reads out the effective allowance. Weeks start on Monday, and the data was exported on 4 October 2026.

Period	Output tokens per week	Total tokens per week	API-equivalent value per week
August (weeks of 3, 10, 17, 24 Aug)	6.68M	1,887.3M	$1,112
31 Aug to 20 Sep (3 weeks)	5.16M	3,591.8M	$2,386
21 Sep to 4 Oct (2 weeks, Opus 5.5)	2.44M	977.5M	$420

21 Sep to 4 Oct against	Output tokens	Total tokens	API-equivalent value
August average	-63%	-48%	-62%
31 Aug to 20 Sep	-53%	-73%	-82%

Opus 5.5 first appears in our logs in the week of 21 September and replaces Opus 5. That timing stands out because Opus 5.5 is cheaper per token at list prices, so a limit proportional to API value would allow more tokens after the switch, not fewer.

Week of (Monday)	Output tokens	Total tokens	API-equivalent value	Opus share of value
3 Aug	8.93M	1,966.3M	$1,004.13	61.6%
10 Aug	3.92M	1,376.9M	$615.28	65.4%
17 Aug	11.50M	3,203.5M	$2,027.66	76.9%
24 Aug	2.38M	1,002.5M	$799.94	81%
31 Aug	5.31M	2,629.4M	$2,071.39	70.3%
7 Sep	5.47M	5,107.8M	$3,205.41	83.5%
14 Sep	4.70M	3,038.1M	$1,882.33	90.6%
21 Sep (Opus 5.5 first seen)	3.91M	1,645.7M	$727.41	87.8%
28 Sep (in progress)	0.97M	309.2M	$113.11	85.1%

Two cautions. The week of 21 September ($727) sits inside the August range ($615 to $2,028), so the fall rests on the averages and on the last week, which was still in progress at export. That week alone was still down 23% in output tokens and 60% in total tokens against the average of the weeks of 7 and 14 September.

How did we measure it?

• Export. ccusage reads Claude Code's local session logs. Its weekly guide lists Sunday as the default start day, so a Monday start needs --start-of-week.
• Weeks versus Anthropic's reset. Anthropic resets the weekly limit at a fixed day and time assigned to each account, not Monday midnight. A capped week on its clock straddles two Monday weeks, so single weeks look high or low. We compare averages over two to four weeks.
• API-equivalent value. ccusage multiplies token counts by list prices. We pay a flat subscription, so this is not money spent. The week of 21 September reproduces exactly at Opus 5.5 list prices ($4 input, $8 one-hour cache write, $0.20 cache read, $20 output per million tokens), and Opus 5 weeks at $5, $10, $0.50 and $25.
• Cache reads and output tokens. Claude Code re-sends the whole conversation on every request, so cached reads make up almost every token and output tokens are a fraction of one percent of the total. Output tokens, thinking included, are still a useful second metric because they ignore context length and cache lifetime. Value and output tokens both fell more than 60% against August, so the finding does not hinge on how cache reads are counted.
• Noise and the last week. Even in August, capped weeks ranged from about $615 to $2,028 of value, and Sonnet-heavy weeks show less. The week of 28 September was still in progress at export. Read the averages and treat single weeks with care.
• Coverage. ccusage sees only logs on disk. Usage on claude.ai, in Cowork or on another computer draws on the same limit and would be missing. Claude Code also deletes transcripts after 30 days by default, which would shrink older weeks and understate the drop.

What has Anthropic published about weekly limits?

We checked Anthropic's pages on 4 October 2026.

Official pages

• The Max plan help page says Max 5x gives five times Pro's per-session usage and that the weekly limit "applies across all models" and resets at a fixed time assigned to your account, with other caps possible at Anthropic's discretion. It gives no weekly allowance in numbers. The usage and length limits article names conversation length, features, model and effort level as drivers, lists no model weights and does not mention cached tokens.
• The Fable models page (updated 2 September 2026) says Fable uses your weekly limits faster than other models, with up to 50% of the weekly limit usable on Fable on Max. Our logs include Fable.
• The Claude Code cost documentation says a long session "still draws usage for the whole conversation" even when history is read from the cache. Cached history counts in some way, and no weight is published.
• The Opus 5.5 launch post (22 September 2026) prices Opus 5.5 at $4 input and $20 output per million tokens and claims 40% lower cost than Opus 5 on typical workloads. For plans it says only that Anthropic is "increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans", with no figure. It is silent on weekly limits and on how Opus 5.5 counts against plan limits.
• BleepingComputer reported on 29 August 2026 that Anthropic would replace its temporary 50% boost to Claude Code weekly limits, running since May, with a permanent 25% increase over the original level from 14 September. Anthropic then said that against the boosted limits this "works out to a 17% reduction in weekly limits on Claude Code". We could not open the original post on X (HTTP 402), so we rely on that quotation. Our August weeks ran under the boost, and the week of 14 September ran mostly under the permanent level yet still reached $1,882.

What other users report

• GitHub issue #99204 (3 October 2026): a Max 5x user on Opus 5.5 at medium effort reports the weekly bar at 33% after about two days. Open, no staff reply visible when we read it.
• GitHub issue #92201 (4 September 2026, before Opus 5.5): a reporter on Opus 5 with a 1M context measured weighted tokens per day falling three to four times while limit consumption rose. Open, no staff reply visible.
• GitHub issue #97467 (26 September 2026): a user who upgraded from Pro to Max 5x says support acknowledged the Max 5x multiplier was not being applied. One report, still open.
• A SmartScope overview (1 October 2026) finds mixed posts on X about Opus 5.5 limits.

As of 4 October 2026, Anthropic has published a 17% cut to Claude Code weekly limits and a rise in five-hour limits. We found nothing from Anthropic on weekly limits changing with Opus 5.5 or on how Opus 5.5 counts against plan limits, and no statement that nothing changed.

Why might the weekly allowance have fallen?

These are possibilities and none is a finding. Usage logs cannot separate them, because Anthropic does not publish how it weights tokens.

1. A tighter weekly allowance. The 17% step on 14 September is documented, and the Max page reserves discretion over weekly caps. A 17% cut alone cannot explain a fall of 48% to 82%. A further change could.
2. Heavier weighting for Opus 5.5 than its API price implies. At list prices Opus 5.5 is cheaper per token than Opus 5: cache reads $0.20 against $0.50 per million, output $20 against $25. Our API-equivalent value fell as well, so a stable limit proportional to value does not match what we saw.
3. Different counting of cached or thinking tokens. The cost docs say thinking cannot be turned off on Opus 5.5, and thinking is billed as output. If cache reads or thinking weigh more against plan limits, the same allowance buys fewer turns, so total and output tokens both fall. Our data cannot tell this apart from a lower allowance.
4. A metering fault on the account. The public tracker has reports of plan multipliers not applying (issue #97467). Our account stayed on Max 5x throughout, but check that the Usage tab in Settings matches the plan you pay for.

Did our workload change?

Capped usage measures the allowance, so it can fall while demand stays flat. Demand did not fall. In the week of 28 September, Claude Code processed about 5.16 billion tokens in total, mostly on non-Anthropic models routed through My Claude Code, against 309.2M on Anthropic models. Model mix is a fair caveat, because a shift toward the pricier model lowers token counts for a fixed allowance. Opus made up 62% to 81% of value in the August weeks and 84% to 91% from 7 September, but the share barely moved across the Opus 5.5 switch (90.6% in the week of 14 September, then 87.8% and 85.1%).

How can you measure your own usage?

1. Run /usage in Claude Code. On Pro and Max it shows plan usage bars, flags long context and cache misses once they reach 10% of recent usage, and toggles 24 hours or 7 days with d and w. Figures come from local history on that machine, so other devices and claude.ai are missing. The Usage tab in Settings on claude.ai shows the weekly bar and next reset.
2. Export weekly tokens with ccusage. Set the start day to your account's reset weekday and add --breakdown for per-model weeks.
3. Keep the history. Our Claude Code session history guide explains the 30-day cleanup. Claude Session Guard keeps a SQLite store of every usage event, de-duplicated by line and by message, with weekly per-model rollups.
4. Note the weekly bar at the same time each day and divide tokens used since reset by the percentage shown. That gives tokens per 1% of the week. A sharp change in that ratio on an unchanged workload is worth reporting to Anthropic support.

npx ccusage@latest weekly --start-of-week monday --json
/plugin marketplace add FiredMosquito831/claude-session-guard
/plugin install claude-session-guard@claude-session-guard

What can you do about it?

• Use Sonnet for routine work. The cost docs say Sonnet handles most coding tasks well and costs less than Opus. Anthropic's model docs list Sonnet 5.5, released 28 September 2026, at $2 input and $10 output per million tokens, against $4 and $20 for Opus 5.5. Switch with /model.
• Keep context lean and cache-friendly. Use /clear between unrelated tasks, /compact with instructions and disable unused MCP servers, because every request carries the full history. Our prompt caching guide explains why a cold cache costs more. On a subscription the cache lasts an hour.
• Route lighter tiers to other models. A gateway such as My Claude Code maps Claude Code's Opus, Sonnet and Haiku tiers to other providers' models, which is how work continued in our capped weeks. Our guide to Claude Code with any model covers the setup. Each provider's terms apply, so check Anthropic's before using a subscription with third-party tools.
• Compare plans. Read Claude subscription vs API for coding and AI coding subscriptions compared before you pay for a higher tier.

How does this connect to what Danube Labs builds?

Danube Labs is a one-person studio in Bucharest that builds LLM tooling daily. My Claude Code routes Claude Code and other coding agents across 57 model providers, and Claude Session Guard keeps the usage history behind the numbers above. If you want help measuring your AI tooling costs or setting up model routing for a small team, book a free intro call.