Claude Sonnet 5.5 – a faster Sonnet at the same API prices
Since version 2.1.284, released on Monday 28 September, you can pick Claude Sonnet 5.5 in Claude Code – the second model in the Claude 5.5 family. Anthropic says Sonnet 5.5 runs more than 30 percent faster than Sonnet 5. It is meant to do best on everyday tasks with a clear goal, on fixing bugs, and on producing polished documents, presentations and spreadsheets. For complex, open-ended work that calls for sound judgement over a long stretch, Opus 5.5 keeps the edge, according to Anthropic.
Opus 5.5 stays the default model; you switch to Sonnet 5.5 with the /model sonnet command. Sonnet 5.5’s default effort level is medium, the same as Opus 5.5’s. If you’re on Opus today only because Claude Code made it the default, this is a good week to switch to Sonnet and see whether you notice any difference in your everyday tasks at all.
API prices haven’t changed: $2 per million input tokens and $10 per million output tokens, the same as for Sonnet 5. Anthropic says that in its tests the same task costs up to 30 percent less on Sonnet 5.5 than on its predecessor, because the new model needs fewer tokens. How much of the Pro or Max limit Sonnet 5.5 uses compared with Opus 5.5, Anthropic hasn’t said.
Our tip Haiku, Sonnet, Opus, Fable – which model for what explains how to tell when Sonnet is no longer enough and it’s time to switch to Opus.
Mods – plugins that change how Claude Code works and looks
Since version 2.1.287, released on Thursday 1 October, Claude Code supports mods. A mod is a plugin that changes how Claude Code looks and works: as Anthropic describes it, a mod is made of small TypeScript functions that can rewrite a prompt, add something to the interface, replace a built-in feature or add an entirely new one. Anthropic started with its own: the built-in /diff command is now a mod, so you can switch it off in /plugin, on the Installed tab. Mods work in the terminal and in the Code tab of the Claude desktop app. In the VS Code extension a mod can change how Claude Code behaves, but nothing it draws will show up there.
You don’t need to know how to program to have a mod of your own. Describe in plain words what the mod should do, and Claude Code writes it. A mod that Claude Code writes works only in the session it was created in – if you want to keep it, ask Claude Code to copy it out of that session’s folder. A request for a mod can be as simple as this:
Make a mod that shows the name of the folder I'm working in above the message box.A mod runs with your permissions, and nothing walls it off from the rest of your computer. It can read and write your files, run programs, connect to the internet and use the model on your plan. It can even approve a tool call on your behalf before Claude Code gets the chance to ask you. The one thing it can’t change is the text of the permission prompt. So treat mods from other people the way you’d treat any program you install: only take them from authors you trust.
Ultracode in /effort now works at every effort level
Ultracode is a Claude Code mode in which Claude organises the work on bigger tasks itself, using Dynamic workflows. Before version 2.1.284 from Monday 28 September, Ultracode was the highest setting on the slider in /effort: it switched on the xhigh level, and picking any other level turned it off. Now it’s a separate toggle that stays on at every effort level. In the terminal you toggle it in /effort with the Tab key or with the /effort ultracode command; in VS Code, with the Ultracode toggle below the Effort slider. Ultracode works on every paid plan, but on Pro you first have to turn on the Dynamic workflows option in /config.
Ultracode is a mode for one big task you’re happy to spend a good chunk of your limit on. The documentation says plainly that with Ultracode on, every request in the session uses more tokens and takes longer, so the session reaches the five-hour or the weekly limit sooner. The /effort ultracode command turns it on only until the current session ends.
Response bookmarks and a Run in background button in the VS Code extension
Since version 2.1.286 from Wednesday 30 September, the Claude Code extension for VS Code lets you bookmark Claude’s responses. The buttons here are described as they looked in early October, and Anthropic changes this extension almost every week. Hover over a response and click Bookmark response. Bookmarked responses go to the Bookmarks panel, which you open with the bookmark icon at the top of the Claude Code panel or with the /bookmarks command. It’s the place for a response you keep coming back to throughout a conversation – the work plan you agreed on, say, or the text of a client proposal you’re still refining together.
The Run in background button arrived in version 2.1.287 on 1 October. When Claude is waiting for a command’s result or for a subagent to finish, and it’s taking longer than you’d like, click Run in background below that command’s or subagent’s entry in the conversation. The button appears once a command has been running for about two seconds, or straight away when a subagent starts work. Claude then stops waiting and gets on with the rest of the task, and when the background work is done, Claude gets a notification.
Since version 2.1.286, the Stop button and the Esc key only end the current turn – the response Claude is working on right now. Subagents running in the background carry on, and you stop them one at a time from the agent map that the /tasks command opens. Since version 2.1.284, the extension can also show the time above every message and, where a new day begins, a line with the date. The option is off by default; you turn it on in the VS Code settings, where it’s called Claude Code: Show Message Timestamps.
/mcp reconnect all in the terminal – one command for MCP servers that have disconnected
Since version 2.1.284, Claude Code in the terminal has a /mcp reconnect all command. It retries the connection to every MCP server that has disconnected or is waiting for you to log in. An MCP server connects Claude Code to an outside service such as a calendar, a mailbox or a document store.
If you ask Claude Code for a list of your meetings from the calendar and Claude Code replies that it can’t access your calendar, there’s no need to work out which connection failed: type /mcp reconnect all and repeat the request.
If you don’t yet know how an MCP server differs from a connector and how to add one, start with our tip Connector or MCP: what actually makes them different.
To finish: six maths papers written together with the Muse Spark model
On Friday 2 October, Meta published six mathematics papers that mathematicians wrote together with its Muse Spark model. Five of them answer questions that had been open until now. The researchers used versions 1.1 and 1.2 in Thinking Mode through the ordinary Meta AI chat, with no special research setup – in other words, much the way you chat with Claude.
People led the work. One team of mathematicians directed the research and developed the arguments with Muse Spark, a second team then checked the first team’s results, and each paper marks which parts the researchers wrote and which the AI did. In a group theory paper, a single counterexample was enough to settle the question – Meta compares it to the black swan that disproves the claim that all swans are white. The counterexample turned up in a group of 384 elements: Muse Spark wrote the program that found it in GAP, a system for mathematical computation, and the mathematicians checked the result and finished the proof.
The second team did its checking within the same collaboration with Meta, and Meta’s post says nothing about peer review in journals, so for now we’re reading the company’s own account of its own work. Meta also acknowledges that teams outside the company solved some of the same problems independently, by other methods. We can’t help wondering how many more open questions are only waiting for someone to find the right exception.