Anthropic announced last week that it is “permanently raising” Claude Code weekly limits by 25 percent, effective September 14. What the company did not lead with: for every developer using Claude Code today, that is a 17 percent capacity cut. Developers noticed within hours, attached a Community Note to the announcement, and forced Anthropic to delete and repost with a clarification that admitted the reduction. The math is straightforward. The messaging choice was not.
The Math, Plainly #
Since May 13, Claude Code has been running a promotional 50 percent weekly limit boost for Pro, Max, Team, and seat-based Enterprise plans. That promotional level — the 150 — is what every developer has been using for the past four months. The new “permanent” level landing September 14 is 125 percent of the original 2025 baseline.
| Measurement | Value | Change |
|---|---|---|
| Original 2025 baseline | 100 | — |
| Summer promo (since May 13) | 150 | +50% |
| New permanent (Sep 14 onward) | 125 | +25% vs. baseline / −17% vs. today |
From 2025 baseline to September 14: up 25 percent. From today to September 14: down 17 percent. Both statements are arithmetically correct. Anthropic led with the first one. Developers were living with the second.
A Promotion That Lasted Four Months Is No Longer a Promotion #
The May 13 boost was originally billed as temporary, running through July 13. Anthropic extended it four times. Three of those extensions were granted after the prior deadline had already passed. By the fourth month, developers had done what developers do: automated workflows around the available capacity. Sprint planning, context budgets, team tooling — all of it calibrated to 150, not 100.
When a temporary promotion survives four extension cycles spanning four months, it stops being a promotion. It becomes the product. Ending it now is a cut, regardless of what number you measure from.
The Community Note and the Deleted Post #
The original announcement on X framed the change as a raise, with the −17 percent acknowledgment tucked into a threaded reply. The developer community caught the gap within hours and appended a Community Note to the original post warning readers about the actual net reduction. Faced with that note gaining traction, Anthropic deleted the original and reposted a clarified version that led with the 17 percent reduction.
The deletion did not go unnoticed. If the original framing was honest, there was no need to delete it. The decision to pull the post rather than correct it in place added a credibility problem on top of a communication problem.
Compute Is Expensive — Anthropic Should Just Say That #
Anthropic’s stated reason for the change is “long-term sustainability.” That is a real constraint. Inference at subscription price points is financially difficult, and the May 2026 5-hour limit doubling was explicitly tied to new SpaceX Colossus 1 datacenter capacity — 300 megawatts, over 220,000 NVIDIA GPUs. No equivalent capacity announcement backed the 50 percent weekly boost. It was likely always going to be temporary.
Anthropic is not the only company navigating this. OpenAI restored a strict 5-hour Codex cap for ChatGPT Plus on August 25, reversing six weeks of uncapped access for the same reason: compute load management. The pattern is consistent across the industry — temporary boosts used for user acquisition, then walked back when the economics bite. That is not a scandal. It is a structural reality of selling AI inference at flat subscription rates. It just needs to be said plainly instead of dressed up as good news.
What Developers Should Do Before September 14 #
If you are on Pro, Max, Team, or seat-based Enterprise: you have one week before weekly capacity drops 17 percent from today’s level. Audit your current weekly usage patterns, identify your highest-token workflows, and consider routing intensive tasks to the Claude API directly, which carries no weekly cap and charges per token. For teams running heavy agentic workloads, diversifying across Cursor, GitHub Copilot, or Gemini Code Assist is worth revisiting. Anthropic did promise forthcoming UI improvements for “better visibility and control” over weekly quotas. Until those ship, you are flying mostly blind — no per-plan token counts have been published, and independent verification is not possible.
The Bigger Issue Is Transparency #
The 25 percent increase is real. The 17 percent reduction is real. Both numbers coexist, and Anthropic’s initial choice of which one to lead with tells you something about how the company handles mixed news. The deleted post and forced repost did more damage to trust than the limits change itself. Developers can absorb a capacity reduction. What they cannot absorb is the suspicion that they need to read the fine print on every announcement. That is a relationship problem, and it is harder to fix than a compute constraint.