Every writeup of Anthropic’s Claude Code limit change has argued about which number is honest, the 25% increase or the 17% cut. Both are. Nobody has bothered with the part that actually affects you, which is that the drop lands on September 14, you are sitting at the high ceiling right now, and Anthropic shipped two things this week that decide how much that drop hurts. That is today’s roundup: one deadline, two tools that arrived just in time to matter, one set of vendor numbers with nothing behind them, and a survey that quietly undercuts the story everyone is telling about it.
Your Claude Code Ceiling Drops in Seven Days
Here is the arithmetic, because the coverage keeps giving you one half of it. Call the pre-promotion weekly limit 100. A temporary 50% boost has had you at 150, and that boost runs through September 13. From September 14 the permanent standard weekly limit sits at 125 for Pro, Max, Team and seat-based Enterprise plans.
So Anthropic’s “we’re permanently raising standard weekly limits by 25%” is true against the baseline. And the number in its own follow-up post the same day is also true:
Compared to today, this works out to a 17% reduction in weekly limits on Claude Code.
Both statements came from Anthropic on August 29. It led with the flattering one, and Bleeping Computer reported that day that the company then deleted the original thread and posted a clarification in which it plainly admits the reduction.
My read: the 25% is not a lie, it is a choice about which baseline to headline, and choosing the baseline your customer is not standing on is spin. Anthropic has been unusually straight with builders on pricing, including the 75% cache read cut earlier this month, which is exactly why this one grates. You do not need the framing. Give me the delta from today and let me plan.
What to actually do with the seven days you have left:
- Run your heaviest recurring jobs now, while you are still at 150.
- Measure a normal week before the 14th, so you have a real before-number instead of a vibe.
- Move anything batch-shaped and non-interactive onto the API, where you are paying per token instead of drawing down a weekly ceiling.
Anthropic Shipped the Diagnostic Two Days Before You Need It
Claude Code v2.1.261 landed on September 5 with /skill-doctor, which lists the skills a session loaded but never used, and what each one cost you in context. Every loaded skill occupies space on your requests whether it fires or not, so on a fat setup this is not a rounding error.
I checked mine while writing this. This project carries 36 skills of its own and there are 75 more installed globally, which is 111 skill definitions eligible to load into a session that will typically use two or three of them. I have been running that setup for months without ever having a number for what the other 108 cost me. Now there is a command that prints it. The same release also raised inline command and background task output limits to 128K characters, which cuts the amount of work that gets truncated and re-run.
The timing here is either luck or good product sense, and it does not matter which. A context-waste diagnostic is a nice-to-have in a week where your ceiling is going up. It is a lever in a week where it is coming down 17%. Prune before the 14th, not after.
Agents Became Terraform on Wednesday
The quieter release is the one I think changes more. Version 1.30.0 of the ant CLI added ant apply, which creates and updates agents, environments, skills, memory stores and deployments from files in your repository. You describe each resource as Markdown, YAML or JSON, run the command, approve the plan it prints, and commit the claude-lock.json it writes so later runs reconcile the same resources instead of spawning duplicates.
That is Terraform, applied to agents. The lockfile tracks resource IDs and hashes, so drift is detectable rather than something you discover when production behaves oddly. In CI you run it with --dry-run on pull requests to print the plan for reviewers and --yes on the default branch after merge.
Anyone who has managed agents by clicking through a console knows what this fixes. Configuration that lives in a web UI has no history, no review, no rollback, and no answer to “who changed this and when”. Moving it into the repo puts agent config through the same review as the code it runs beside. I would rather have this than most of the model releases from the past month, and it got a fraction of the attention because it does not have a benchmark score attached.
Thirty-Five Percent Larger Carts, According to Nobody
Anthropic open-sourced its commerce agents blueprint on September 2 under Apache 2.0, covering a shopping agent and a merchant agent across retail, travel, telecom and entertainment. The engineering is honest about its own limits: the shopping agent hands the cart off to the merchant’s existing checkout rather than processing payment, and the merchant agent only proposes price, restock and campaign changes for a human to approve.
Then there is the sentence doing the selling. Anthropic’s own post says retailers running shopping agents on Claude see “carts up to 35% larger and shoppers 60% more likely to complete a purchase”. No sample size. No methodology. No named study. No indication of what those merchants were doing before, which for a conversion claim is the entire question.
I want to be precise about what I am objecting to, because the code itself looks genuinely useful and Apache 2.0 is a real gift. The objection is that a company selling infrastructure to engineers dropped two conversion statistics with less rigor than those engineers would accept in their own standup. “Up to 35%” is a phrase from a mattress ad. Ship the blueprint, keep the numbers, and either publish the study or cut the sentence.
A Third of Companies Stopped Buying Software. The P&L Did Not Notice.
McKinsey’s State of AI global survey found that 32% of organizations passed on buying a software product because they could build the functionality in-house with agentic coding tools. Among the roughly 6% of respondents who attribute at least 5% of EBIT to AI, close to half are skipping purchases, against 31% of everyone else. The survey ran May 4 to June 8 across 1,719 respondents in 97 countries, weighted by each country’s share of global GDP.
That finding has been everywhere this week as “build versus buy just flipped”. Read one line further. The share of organizations attributing any earnings impact to AI sat flat year over year at 37%, and the high-performer group held at about 6%, exactly where it was in 2025.
So a third of these companies replaced a vendor invoice with an engineering commitment, and the aggregate P&L is unchanged. That is not a failure, it is a timing mismatch, but it does mean the honest framing is a cost transfer rather than a saving. You stopped paying a SaaS bill and started paying to run, patch, secure and staff the thing you built, and the run cost is a ceiling that vendors like Anthropic reprice on their schedule. Which is where this roundup started, and why the 14th is not a small story. If you built in-house on agentic coding tools this year, your inputs are already priced closer together than they were, and your capacity is about to be worth 17% less than it is today.
What I Am Watching
The deadline is the only item here with a date on it, so treat the other four as preparation for it. Prune the skills, move batch work off the weekly ceiling, and put your agent config in the repo while you have slack to do it carefully. On the commerce blueprint, I will believe the conversion numbers when someone publishes a denominator. And on build versus buy, the number worth tracking is not how many companies stopped buying software this year. It is how many are still running what they built in eighteen months.