Claude Opus 5.5 is the new default model, on the Pro plan too
Claude Opus 5.5 came out on Tuesday 22 September as the first model of the new Claude 5.5 family. Anthropic says it works at the level of Claude Fable 5.1 on most tasks, responds more than 30 percent faster than Opus 5 and writes more clearly, with the most important information first. According to the announcement, it is far less likely than earlier models to take actions that are hard to undo or to step outside the limits it was given, and it resists prompt injection – instructions slipped into text it reads – better than Opus 5.
In Claude Code version 2.1.280, released the same day, Opus 5.5 became the default model on Max, where Opus 5 had held that place, and on Pro, where Sonnet had been the default. If your settings pin a different model, Claude Code keeps using it. You change the model as before, with the /model command.
Opus 5.5 costs 4 dollars per million input tokens and 20 dollars per million output tokens, twice as much as Sonnet 5. Opus 5 costs 5 and 25 dollars, two and a half times the price of Sonnet. The price gap between Opus and Sonnet is therefore smaller than it was, which helps anyone who pays for Opus through the API. These are rates for tokens billed through the API – the price list says nothing about how much of your Pro or Max limit each model uses up.
Changing the effort level on Opus 5.5 no longer wipes the cache
Claude Code keeps the part of the conversation the model has already read in the model’s prompt cache, so the model doesn’t read everything from scratch for its next reply. On most models, changing the effort level mid-session invalidates that cache, and the next request reads the whole conversation history again from the start. That is why Claude Code, while the cache is fresh, asks you to confirm the change first. On Opus 5.5 a change of effort level leaves the cache alone, and Claude Code switches the level without asking. Opus 5.5 is the second model in Claude Code that works this way – Fable 5.1 has done so since version 2.1.260, from Thursday 3 September.
Opus 5.5 starts at medium, one step lower than Opus 5. Anthropic reports that in its tests Opus 5.5 on medium does at least as well as Opus 5 on high. We recommend staying at the default level and raising effort only for a harder stretch of work. You can now do that halfway through a conversation with the /effort command and go back down afterwards, without Claude Code reading everything again each time.
Switching models still rebuilds the cache from scratch, because each model has its own.
We explain why switching models mid-conversation costs more than it seems in the tip Haiku, Sonnet, Opus, Fable – which model for what.
A higher five-hour limit, and a limit reset you can use when you choose
Together with Opus 5.5, Anthropic announced a higher five-hour limit on the Pro and Max plans, though it didn’t say by how much. The weekly limit changed separately, and we covered that on 14 September. Subscribers also got a limit reset they can use whenever they choose. A reset restores the full five-hour or weekly limit at once – which of the two it covers is shown in the reset’s description in your account. A reset can’t be undone, and it is lost if you move to a lower plan or cancel your subscription before using it.
You’ll find the Reset for free button in claude.ai in the browser or in the Claude desktop app, under Settings → Usage, in the Resets section. If a reset has an expiry date, you’ll see it in the same place. In claude.ai and in the desktop app, the same button also appears in the message telling you the limit has run out. Anthropic rearranges the claude.ai settings fairly often, so the button may move elsewhere over time.
Claude Code in the terminal and in VS Code has no reset button, but the limit is shared across your whole account, so after a reset your limit is full there too. A reset is best kept for the day your limit runs out in the middle of important work, because a reset used while half your limit is still left only refills the half you’ve used.
The y and n keys no longer confirm and close Claude Code dialogs in the terminal
Pressing y on its own no longer confirms a dialog in Claude Code in the terminal, and n no longer closes one – that has been the case since version 2.1.280. Until now both letters were bound to those actions by default, so a letter pressed by accident could approve something. Approving anything with a single letter was risky, so we’re relieved to see this fix. You now confirm a dialog with Enter and cancel it with Esc. The exception is a dialog that shows y and n as its own keys – that one still responds to them.
You can bring the old binding back in your keyboard shortcuts file, keybindings.json – just ask Claude Code to do it. If that file was created with the /keybindings command before version 2.1.280, both old bindings are written into it and keep working until you ask Claude Code to remove those two lines.
The new maxProseWidth setting narrows reply text in a wide terminal window
The maxProseWidth setting limits how wide the text in Claude Code’s replies runs, and it works from version 2.1.282, released on Thursday 24 September. It helps when your terminal window fills a whole wide monitor: reply lines then run from one edge to the other, and your eye struggles to land on the start of the next one. With this setting, paragraphs, headings and lists wrap at the width you give, while tables and code blocks still use the full window.
The width is given in terminal columns, which is roughly characters per line, and it can’t be less than 40. The setting is absent by default, so text reaches the edge of the window. To turn it on, ask Claude Code – the 100 below is just a sample value, and the smaller the number, the narrower the text:
Set maxProseWidth to 100 in my Claude Code settings.Ctrl+Enter delivers your message at once and no longer stops background work
Claude Code lets you add a message while it’s working on a task. When you press Enter, the message goes into a queue and Claude gets it once the current step is done. Since version 2.1.275 of 17 September there has also been a Ctrl+Enter shortcut that sends the waiting messages right away – except that it always interrupted whatever Claude was doing at the time.
Since version 2.1.281, from Wednesday 23 September, Ctrl+Enter no longer interrupts that work if it can carry on in the background – a longer command, say, or a subagent that is running. That work keeps going, and Claude reads your message straight away. The shortcut now interrupts only when Claude is in the middle of writing a reply or doing something that can’t move to the background. It comes in handy when, for example, Claude Code is tidying up a folder of documents and you remember that the folder with your contract templates has to stay untouched.
In some terminals Ctrl+Enter behaves like a plain Enter – you can tell because the message stays in the queue instead of reaching the AI right away. In that case Ctrl+X followed by Ctrl+S does the same job, and that pair works in every terminal.
The VS Code extension: a Status window, conversation export and a skill switch
The Claude Code extension for VS Code picked up several new features in version 2.1.280. You open all of them from the command menu, which appears when you click the / button in the chat’s text box or type / there. Each one also has its own command. The extension has been changing every few days lately, so the sections of that menu may still move around.
The Status window in the Customize section – you can also open it with the /status command – shows the Claude Code version, your account, the model and the MCP servers of the current session. It’s where you check which model you are really working on. Export conversation in the Context section, also available as the /export command, copies the conversation as plain text or saves it to a file. If you add a file name straight away, for example /export notes.txt, you skip the choice dialog and go directly to picking the folder the file should be saved in. It’s a good way to keep a record of what you decided in a long conversation before you close it.
The third new feature is about skills. The Customize section has a Slash commands window listing every command and skill, and since version 2.1.280 it shows, for each skill, where it comes from, roughly how many tokens it takes up and whether it is switched on. You change a skill’s state with a click, unless the row is marked locked – as it is, for example, with skills that come from plugins. The new /skills command opens the same window. Look into it when a session fills its context quickly: you can see at once which skills take up the most room. The /plan command is new as well: it switches the conversation into plan mode, and with a task description added it asks Claude Code for a plan straight away.
Anthropic opened a portal for submitting plugins to the Claude directory
On Friday 25 September Anthropic opened a portal through which creators submit plugins to the Claude directory. A plugin is a package in which its creator bundles skills, connections to outside services (MCP connectors), or both. You browse the directory in claude.ai and in the Claude desktop app, on the Customize page, and a plugin added from there also works in Claude Code sessions.
Anyone with a paid Claude plan can submit a plugin, without signing up for a partner programme first. We like that a single person can submit their plugin there too, not only a company. Anthropic reviews each submission, security included, as soon as it’s sent, and once it’s accepted the creator decides when the plugin goes into the directory. Do you have a skill that others in your line of work would use? Package it as a plugin, and you can now submit that plugin there. If you don’t build plugins, you still gain indirectly: over time the directory will fill up with plugins that have passed Anthropic’s review.
Finally: museums look to AI to bring people closer to art
On Wednesday 23 September, The Art Newspaper reported on a study by Remuseum, an independent research project organised by the Crystal Bridges museum in Bentonville, Arkansas. The report grew out of a meeting of representatives from seven American museums, among them MoMA, the Cleveland Museum of Art, the Detroit Institute of Arts and the Frye Art Museum in Seattle. Its authors argue that AI can help museums without the job cuts the sector fears so much, and they don’t gloss over the ethical and environmental risks or the copyright disputes.
The five uses it describes range from fundraising through communication to visitor services. The Frye uses AI to keep a consistent tone in its communication, in line with a style guide written by its staff. The Detroit Institute of Arts connected its databases into what it calls an “institutional brain”, and staff can now pull information out of it more easily. The idea we liked best comes from the Toledo Museum of Art and hasn’t been tried yet. There, AI would work out which visitors stood longest in front of a particular Monet, and the museum would invite twenty of those people to a conversation with the exhibition’s curator.
Stephen Reily, the director of Remuseum, says in the article that the opportunities the report’s authors saw were less about automation and cutting jobs and more about strengthening people’s connection with museums and with art. The public is wary of AI in museums: in another survey the article cites, more than 70 percent of respondents opposed the use of AI in exhibitions. Jamilee Lacy, the Frye’s director, admits that after several days of talking about AI she still found it hard to think of anything better than looking at a painting in person.