Claude Code will not tell you that your weekly limit has changed – you notice it only once your tokens run out faster than usual. Beyond that, Claude Code picked up a few small things you see straight away: it stopped printing API keys on the screen, and it reloads plugins without a command of its own.
The weekly limit in Claude Code changes today
Anthropic announced at the end of August that from 14 September it is permanently raising the standard weekly limit in Claude Code by 25 percent. The same announcement carried a second piece of news: the 50 percent increase announced on 13 May ended on 13 September. It had an end date from the start – 13 July at first – but Anthropic extended the increase again and again, so over four months we all came to treat it as the norm. So today’s return to a lower limit feels to us less like a promotion running out than like losing something that had become permanent. From today both decisions apply at once. Compared with last week the limit is 17 percent lower. Compared with where it stood before May, it is 25 percent higher. Those 17 percent come from Anthropic itself. It deleted the first version of the announcement and published it again, adding how far the limit drops against what applied until yesterday.
The change covers the Pro and Max plans, and in companies also Team and Enterprise billed per seat. It applies to the weekly limit alone – the five-hour one stays as it was. How much of the weekly limit you have left is shown by the /usage command.
Anthropic gave no reason for the change. Outside that post on X the company described it nowhere at all: we searched the Claude help centre and the company blog. If you plan your working week around the limit, look into /usage earlier today than you usually would.
A plugin works as soon as you close the /plugin panel
Since version 2.1.268, released on Thursday 10 September, Claude Code reloads plugins the moment you close the /plugin panel. You install, enable, disable and uninstall a plugin without leaving the conversation and without typing the /reload-plugins command. Until then, changes made in that panel waited for you to type it. It was easy to miss, because the panel closed exactly as it does today, and a freshly installed plugin simply did nothing.
Sometimes the reload does not happen after all: Claude Code warns you that the changes have not taken effect. You push them through with /reload-plugins --force. Plain /reload-plugins, without --force, still has one job left: an update notice for a plugin sometimes asks you outright to reload.
Claude Code no longer shows API keys next to MCP servers
That same version of Claude Code, 2.1.268, stopped showing keys and passwords from MCP server configuration on the screen. A token for a service goes into that configuration either directly or as ${VAR}. For that notation Claude Code substitutes the real value, and the real value is what could show up before: in the server details in /mcp and /plugin, in the output of claude mcp list and claude mcp get, and in the error message when logging in to a server. The fix reached plugin error messages too: they showed a token or a password written into the address of the repository the plugin comes from.
The fix covers only what Claude Code puts on the screen. And more people see that screen than you would think: the output of claude mcp list gets pasted into a bug report, and the server details are on view in a meeting when you share your desktop. If one of your keys did show up that way, generate it again in the panel of the service it belongs to.
You create the token for an MCP server configuration beforehand, in the service’s own panel, and some services ask you to simply log in instead. We describe both routes in the tip Connector or MCP: what actually makes them different.
The new maxEffortLevel setting caps the effort level
Claude Code gained the maxEffortLevel setting in version 2.1.267, released on Wednesday 9 September. The effort level decides how much the model thinks about each step: at a lower one it works faster and more cheaply, at a higher one it reasons more deeply on harder problems. maxEffortLevel marks the upper bound, and every higher level gets cut down to it – the one from the /effort command, the one that comes with the model you pick in /model, and the one written permanently into the effort field of a skill or a subagent, if you use those.
There are five values: low, medium, high, xhigh and max, where max means no ceiling at all. By default the setting is not in your configuration, so Claude Code cuts nothing. maxEffortLevel goes into the ordinary settings file – it is enough to ask Claude Code to put it there.
A ceiling is worth setting when the level is fixed in a skill or an agent written a long time ago, and with subagents that always ask for max. For a lower level the documentation of this setting promises only faster and cheaper work. About the weekly limit it says nothing. So we would not assume in advance that the ceiling will make your weekly limit last longer.
The VS Code extension: a permission rules window and a subagent counter
The Claude Code extension for VS Code changed in several places, most of all in the area where you type. Anthropic has been rebuilding the bottom of the chat window piece by piece over the past two weeks, so your version of the extension may show a different set of elements down there than ours does. The name of the model you are working on has been sitting there since version 2.1.257 of 1 September. Next to it you see the effort level you chose – the same level that maxEffortLevel has been capping since 9 September. The effort level joined it in version 2.1.260 of 3 September, so if you are seeing either of them for the first time: they are not new this week.
Since version 2.1.269, released on Friday 11 September, permission rules work in VS Code too – the same ones the /permissions command has long shown in the terminal. You find them under Permissions in the Customize section, in three groups: Allow, Ask and Deny. You add a rule of your own on the spot, and remove one you saved earlier.
That same version 2.1.269 also put a count of the subagents working in this conversation at the bottom of the chat window – the separate threads Claude Code runs alongside the main one. The counter looks like 2 agents, and the dot beside it shows whether a subagent is working right now or waiting for your approval. Clicking the counter opens a map with the state of each of them. With no subagents running, there is no counter.
The conversation list changed in version 2.1.265, of Tuesday 8 September. A conversation in which nothing happened for 14 days lands in the Archived sessions group at the bottom of the list, with no help from you. Nothing in it is lost – you can bring it back with Unarchive session, and you set the length of that break in the Archive Inactive Sessions option.
Our tip Manual, accept edits or auto: how much to let Claude Code do goes through which answer to a permission question saves a rule for good, and how that works in each of the permission modes.
To close: the head of Anthropic writes that AI has to slow down
Dario Amodei published an essay this week, “We Must Pace the Frontier”. In it he argues that the companies building the most powerful models should deliberately slow the growth of what those models can do, so that the work on safety can keep up. The man writing this runs a company that makes its money from that pace.
Anthropic commits to the first of the three proposed steps straight away: a standing invitation for external reviewers who become part of the company’s own team – they get desks, access badges and company laptops. Their job is to check whether the company keeps its own safety commitments, and they will be free to publish what they find without Anthropic’s approval. The second step is a voluntary agreement between AI companies in democratic countries: shared safety standards and a shared limit on how fast model capabilities grow. The third is talks between democratic governments and authoritarian ones – as far as those are possible at all.
Sam Altman and Elon Musk backed Amodei’s appeal publicly, and Altman said that OpenAI would let independent reviewers in as well. We report it as a curiosity, because for now it all remains at the level of declarations: the reviewers are yet to be invited, and Amodei himself warns that nobody will feel the slowdown, because progress will still look fast. We will see how much of this backing turns into something outsiders can check.