Leadership2 distinct publishers3 min readPublished
The 50% promotion that has run since May ends on September 14, and what replaces it sits below today's ceiling, so teams that planned September work around summer capacity are re-planning against a smaller weekly cap.
The Board Room · Leadership desk

Compiled by The Board RoomSomething wrong?How this is made
Fifty index points of weekly capacity went on in May; twenty-five come off on September 14 and twenty-five stay [5][14]. That is half the promotion folded into the standard allowance and half retired [15]. The floor a team can plan on into next year therefore sits above April's, while the ceiling it has worked against since mid-May comes down [5].
Capacity planning runs off a different figure than the one in the announcement. A team that wants to hold this summer's throughput past the 14th has to replace 25 index points, and measured against the new permanent level that is a 20% top-up rather than a 17% one [16]. Anthropic has published no absolute token or prompt counts for any plan, so that fifth cannot be turned into a purchase-order line; it can only be read off a team's own consumption in the weeks either side of the change [7].
How much it hurts depends on the shape of the week. Anthropic's plan multiples describe per-session usage, while a separate weekly limit runs across models and resets at a fixed time assigned to each account, which is why a subscriber can be out of weekly capacity with a fresh five-hour session available [19]. Karl Kahn's proposed class action, filed on June 14 in the Northern District of California, alleges that one full five-hour session consumed 15% of his weekly allotment after an April upgrade [17]. Rebase that allegation onto the new cap and the same session becomes about 18% of the week, roughly 5.6 full sessions where there had been 6.7 [18]. No class has been certified, and Anthropic declined to comment [17].
A skeptic on the finance side would say a permanent 25% above the pre-May baseline is a better asset than a promotion that can lapse at any renewal date, and as a description of the trade that is correct [2]. The answer is that the two halves of it are paid by different people. Removing the expiry date is worth something to whoever signs the renewal and nothing to the engineer who committed September work against the summer ceiling, and the promotion's deadline had already been extended several times, which is what teaches a team to plan on it [14].
The paperwork has not caught up with the policy. The Help Center entry still describes the boost as running to 11:59 PM Pacific on August 31 before reverting to the standard cap, while the ClaudeDevs developer account puts the changeover at September 14 [8], and no exact switchover time or regional sequence has been given [9]. The scope question is sharper for administrators: the announcement names Claude Code weekly limits on the Pro, Max, Team and seat-based Enterprise plans [10], and says nothing about the Free and usage-based Enterprise plans the May boost excluded [13].
This week's item is a re-plan. The longer-run item is that per-seat coding capacity is metered in units no buyer can audit, which is what lets a 25% raise and a 17% cut describe the same event [7]. The five-hour session limits doubled on May 6 are untouched by any of this [12], the Reddit complaints circulating are anecdotes rather than a churn count [21], and Anthropic says it is working on changes meant to make customers feel they get more from Claude [23]. Until those arrive, the defensible September plan is the smaller number.
Ranked by verification strength, evidence, and original report placement.
On August 30 Japan time, Anthropic announced it would permanently raise Claude Code's standard weekly limit by 25% starting September 14.
Anthropic said: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code."
On an index where the pre-May standard weekly limit is 100, the temporary 50% boost took the available limit to 150, and the level that becomes permanent on September 14 is 125.
(125 - 150) / 150 comes out to about -16.666%, which rounds to the 17% decrease from current levels that Anthropic cited.
Anthropic has not disclosed the absolute token or prompt counts for each plan; the figures 100, 150 and 125 are index numbers for comparison, not actual usage counts.
The September 14 announcement names the Claude Code weekly limit for the individual Pro and Max plans and the organizational Team and seat-based Enterprise plans.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 1, 2026
1 article · September 1, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
product
Binance gives agents a trading seat, and gives users the permission slip1 distinct publisher
product
A 2x LLM bill is not a bug report: token spend is an observability problem1 distinct publisher
product
OpenAI's plan to hand everyone a coding agent leaves the hard part to the model1 distinct publisher
build
Anthropic's federal ban falls on a record that failed the supply-chain statute1 distinct publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Vendor-sourced but independently reproduced
The spine of this story is arithmetic anyone can check, and two publishers working separately arrive at the same 100 → 150 → 125 and the same 16.7%. What keeps it out of the high range is that every input is Anthropic's: the percentages, the plan list, the dates. XenoSpectrum is explicit that the index numbers stand in for token counts nobody outside the company has seen, and the most concrete usage figure in circulation — 15% of a week per session — sits in an uncertified complaint.
No usage or churn data exists
Nothing here measures uptake or departure. Anthropic publishes no absolute weekly counts, so even the size of what is being withdrawn is expressed as an index; the customer response amounts to two named Reddit accounts, one of whom says he will keep Claude as his primary tool. implicator.ai says outright that these posts cannot establish a trend, and no cancellation figure is tied to September 14. What can be observed is vendor policy, not user behaviour.
Vendor headline overstated; the reporting corrects it
The overstatement belongs to the announcement, not to the coverage. "Permanent 25% increase" is arithmetically true only against a baseline nobody has used since May, and both publishers refuse to launder it. A residual gap remains on the other side: the widely repeated 17% describes limit points, and XenoSpectrum is careful that it does not convert into 17% fewer finished tasks, while restoring current throughput actually calls for a 20% top-up. Precise-sounding percentages here are doing looser work than they appear to.
Everyone quoted has a stake in the framing
Follow who benefits from each number. Anthropic recovers compute by letting 25 index points lapse while describing the day as an increase, and promises unspecified changes to make customers "feel they get more." OpenAI's product lead volunteers, the day after, that Codex's 20x is a weekly figure with no session cap — competitive positioning offered as clarification. Plaintiff's counsel needs the labels to look misleading. And implicator.ai's piece carries its own newsletter pitch mid-article and is an editor-reviewed AI summary, which shapes what gets restated rather than reported.
The what is settled; the when and how much are not
That capacity falls on September 14, and by roughly a sixth, is as solid as a single-vendor fact gets — confirmed in the company's own words and checked twice. Below that, the record wobbles: Anthropic's help page and its developer account gave different end dates, no switchover hour or rollout order exists, previously excluded plans go unmentioned, and the practical effect on delivered work is unmeasured by anyone. Two publishers is thin corroboration when both are reading the same posts.