Claude Code weekly limits drop 17% on September 14
Anthropic is permanently raising Claude Code weekly limits 25% on September 14 — which is 17% less than the temporary boost you have today. Plan the delta.
Claude Code weekly limits are being made permanent, and that is a reduction. Anthropic's developer account announced on X that starting September 14 it will permanently raise standard weekly limits in Claude Code by 25% for Pro, Max, Team, and seat-based Enterprise plans — while the temporary 50% increase stays in place until then. If your team's throughput is calibrated to today, you lose about a sixth of it in two weeks.
What actually happened
The arithmetic is the story. A permanent 25% raise measured against the old baseline is 1.25. The boost running right now is 1.50. Divide and you get 0.833 — roughly 17% less headroom than you have this week, arriving September 14. The headline says increase. Your queue will read it as a cut.
This is the end of a pattern rather than a surprise. Anthropic has been stacking temporary boosts on Claude Code all year and extending them in short increments — Help Net Security tracked one such extension in July that pushed a 50% bump out by six days. Back in May the company doubled Claude Code's five-hour rate limits and removed the peak-hours reduction for Pro and Max, explicitly tied to new capacity: more than 300 megawatts and over 220,000 GPUs coming online at SpaceX's Colossus 1.
Anthropic framed the September change as what it can "sustainably serve going forward," and pointed at future work on usage visibility and control. Translation: the promotional era is over and the permanent number is lower than the promotional one.
Why an agent capacity cut matters for your business
Weekly limits are a staffing decision now. If two engineers and a fleet of subagents are producing at a rate that assumes 1.5x, that rate drops on September 14 and nobody files a ticket about it. Measure your current weekly consumption before the change so you know whether 0.833x is a rounding error or a missed sprint.
Five-hour windows and weekly windows fail differently. Burning a weekly cap on Tuesday is not a slow afternoon, it is a dead week. This is the same failure mode we wrote about when coding-agent token caps started landing at Uber and Tesla: the constraint moves from cost to calendar.
Own the harness, rent the model. The fix is not a second subscription. It is keeping your agent loop, prompts, and evals in your repo so you can route overflow work to a second provider on a bad week — the argument we made for BYOK coding harnesses. A subscription limit should cost you a config change, not a plan.
Key takeaways
- Starting September 14, Claude Code standard weekly limits rise a permanent 25% for Pro, Max, Team, and seat-based Enterprise
- The current temporary 50% increase runs until that date, making the change ~17% less headroom than today
- Anthropic has extended temporary boosts repeatedly through 2026; this replaces them with a lower permanent floor
- May's limit increases were tied to new capacity — 300+ MW and 220,000+ GPUs at SpaceX's Colossus 1
- Measure your team's current weekly burn now, before the baseline moves under you
- Keep the agent loop and evals in your repo so overflow can route to another provider
A vendor's weekly cap should not set your delivery date. We build agent harnesses that live in your repo — your prompts, your evals, your routing — so a limit change is a config value instead of a stalled sprint. See how we keep AI tooling swappable, or tell us what your team burns in a week and we'll size the gap.
Sources: Anthropic (@ClaudeDevs), Anthropic, Help Net Security.
- #claude-code
- #ai-agents
- #developer-tools
- #ai-cost
- #capacity-planning
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
VS Code 1.135: a second model reviews your agent's work
VS Code 1.135 ships Rubber Duck, an experimental second-opinion pass from a complementary model, plus per-turn token accounting. Build the same loop in CI.
Read itOmacom Foundation's $10M: who holds your tool's trademark?
Ten tech founders put $1M each behind a Linux distro's foundation. The interesting line isn't the money — it's that the foundation holds the trademarks.
Read it