Claude Code's 'Permanent 25% Increase' Is Actually 17% Less Than Last Week
Anthropic's new weekly limits for Claude Code took effect on 14 September. They are higher than the spring baseline and lower than what users actually had, because a temporary boost expired the day before.
Anthropic’s new weekly usage limits for Claude Code, its tool for letting the model write and edit code directly in a project, came into force on 14 September. The company announced them as a permanent 25 percent increase for Pro, Max, Team and seat based Enterprise plans. For anyone who was using the tool last week, it is a cut.
The arithmetic is simple once you know the missing piece. Since May, Claude Code had been running a temporary 50 percent boost on weekly limits, and that boost expired on 13 September. So if the original allowance was 100, users had 150 on Saturday and 125 on Monday. That is 25 percent more than the spring baseline, and about 17 percent less than what people actually had in hand. Anthropic deleted its first announcement thread after the framing was criticised and posted a clarification that stated the number plainly: “Compared to today, this works out to a 17% reduction in weekly limits on Claude Code.” The company added that it is working on changes to give users “more visibility and control” of their usage. Free plans and consumption based Enterprise seats are not affected, and the separate five hour session windows on Max plans are unchanged. Anthropic also notes that a weekly allowance does not convert into a fixed number of prompts, because usage depends on how long your conversation is, which model you pick, how many tools it calls and how hard it is told to think.
What is behind this
Every AI subscription is a bet that the average customer will use less than they pay for, and coding agents broke that bet. A chat user sends a question and reads an answer; a coding agent can run for twenty minutes, read fifty files and call tools hundreds of times on a single instruction. Anthropic has reportedly pointed to difficulty securing enough computing capacity to keep the higher ceiling permanent, which fits the wider picture of every large lab signing enormous multi year compute deals right now. The more interesting detail is the communication. Announcing a reduction as an increase, then deleting and correcting it, is the kind of thing that costs more trust than the ten percentage points were worth.
What this means for you: If you pay for Claude Code, plan for roughly a sixth less headroom per week than you had in early September, and notice that the limit is measured in work done rather than messages sent. The cheap habits are the useful ones: start a fresh conversation for a new task instead of dragging a long history along, point the tool at specific files rather than a whole repository, and use a smaller model for routine edits. If you do not use coding agents, this is still a useful signal about where subscription pricing is heading across the industry: flat monthly fees for heavy automated work are proving hard to sustain, and limits are likely to keep moving.
Sources
Google's New Voice Models Keep Talking While They Think
Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, two speech to speech models built for voice assistants. The headline trick is that the slower one reasons and speaks at the same time instead of going silent.
