Anthropic's official developer account announced on August 29 that starting September 14, the standard weekly usage limit for Claude Code will permanently rise 25% across Pro, Max, Team, and seat-based Enterprise plans. Until then, the current 50% temporary boost stays in effect through September 13.
The two percentages only make sense together. Say the original baseline is 100 units: the promotion that started in May bumped that to 150, and after September 14 it settles at a fixed 125. Relative to the baseline, that's a 25% gain. Relative to what people are actually using right now, it's a 17% cut.
A post that got deleted and reposted
The first version of the announcement only stated half the story: a permanent 25% increase. Developers did the arithmetic themselves within hours, and the backlash converged on one point — the good news got announced, the bad news got left out.
Anthropic subsequently deleted the original post and republished it with the missing line added: "Compared to today, this works out to a 17% reduction in weekly limits on Claude Code."
Deleting and reposting is rarely a cheap move for a subscription product's communications. That Anthropic was willing to eat that cost suggests the company judged that being caught out was more damaging than disclosing it upfront.
The limit was never easy to count in the first place
Anthropic's own explainer stresses that the weekly limit isn't a fixed number of conversations. Actual consumption depends on conversation length, which model is selected, how much tool-calling happens, and the effort level in use. Under the same "weekly limit," someone running long-context sessions heavily on Opus and someone making light edits with a smaller model can burn through wildly different amounts.
The direct consequence is that users can't predict their own position in advance. Someone who maxed out the promotional 150 has no clear sense of where they'll land within the new 125 — they'll only find out through the feel of next month's usage. In the same explainer, Anthropic said the changes underway are meant to make people "feel like you're getting more from Claude, while having more visibility and control of your usage." If that visibility had shipped first, the delete-and-repost episode probably wouldn't have happened.
Both companies are pulling their promotional pricing back
There's a related development in the same timeframe. That same week, OpenAI reset usage and billing for Codex and fixed a batch of abnormal consumption issues — including context compaction, memory tasks, goal tasks, automations, subagents, and MCP tools — with the company saying some goal tasks had previously been eating up 15% to 70% of the weekly quota.
The two moves look like they're pointing in opposite directions — one company pulling capacity back, the other handing back capacity that was being wasted — but the underlying pressure is the same: the unit cost of coding agents still hasn't dropped low enough to keep the taps open indefinitely. The models themselves are getting cheaper, but agent usage patterns are getting more expensive — a single task that used to take a few rounds of conversation now runs dozens of tool calls, and the savings on per-unit price get eaten up by call volume. Once the subsidy period ends, the math has to be shown to users again.
From a product-cadence standpoint, this also reads as a signal that subscription AI tools are entering a second phase. In phase one, quota was the growth lever — whoever gave the most captured the most developers. Now that users have moved their workflows in and switching costs are high enough, pricing power is shifting back to the supply side. Going forward, the gap between competitors is less likely to come from who hands out more capacity per week, and more from how much usable output can be squeezed out of the same quota.
Doing the rough math on how it feels: a Max user who currently runs their weekly quota to the limit will effectively lose more than a day's worth of heavy usage per week after September 14. For an individual developer, that's pushing Friday afternoon's refactor into next week. For a team paying per seat, it's roughly one person's worth of capacity evaporating out of every six — something that has to be re-budgeted.
That promotions eventually end isn't in dispute. What actually triggered the backlash was the framing: the same number reads as addition from the baseline and subtraction from the current position, and the announcement only wrote down the addition.
Sources: Anthropic official developer account announcement, BleepingComputer, CocoLoop; the 25%, 50%, and 17% figures were checked against the original baseline, current promotional limit, and new permanent limit respectively.