Claude Code got four changes in one week, and one of them retires a sentence we wrote seven days ago. Outside the terminal, Anthropic handed its top-shelf model to security teams and took the agent-building tools out of beta.
The company also spent last week showing off its own course platform, Claude Academy. The address and the layout are new there; the courses have been running since March.
Your session comes back on its own when the limit resets
Claude Code can now resume work it had to abandon, at the moment your usage limit resets. Until now the sequence went like this: the limit ran out mid-task, the session stopped, and there it sat until you came back and told it to carry on. Version 2.1.234 made that automatic, and if you would rather keep a hand on it, /config has a switch called Continue automatically at usage limit.
We would leave it on if you tend to set Claude Code going and walk away for hours. Bear in mind, though, that what the AI is allowed to do while you are away comes down to your permission mode. Which mode suits which stage of the work is the subject of our tip on permission modes.
A fifth output style: Concise
An output style is the Claude Code setting behind the role, the tone and the shape of an answer. Four of them existed already – Default, Proactive, Explanatory and Learning – and Concise now joins them. Claude opens with the result, drops the preamble and the account of what it just did, and keeps things short. Anthropic says the work itself is done as thoroughly as in the Default style. Ask for an explanation and it arrives in full. Error reports, security warnings and questions about irreversible actions stay complete, which is the right call: those are the last things worth trimming.
Our guess is that the style came out of complaints about how much the fifth-generation models talk. It has been turning up in Claude Code’s issue tracker since July, where one report asked outright for a concise mode and called the answers it was getting walls of inscrutable text. Concise reads like the answer to exactly that.
You pick the style in /config under Output style, and it needs version 2.1.237 or newer. What each of the five does, and the kind of work each one suits, is in our tip on the Claude Code config.
Messaging between sessions works on Windows now
A week ago we wrote that passing messages between Claude Code sessions left native Windows users out. That sentence has expired: version 2.1.239 added Windows, so sessions open across your machines can find each other and write to each other exactly as they do on macOS and on Linux.
Mythos 5 goes to the security teams
Claude Mythos 5 is the same model as Claude Fable 5, the most capable one Anthropic releases widely. A single thing separates them: Fable carries safety classifiers and can decline a request, while Mythos has none of them and declines nothing.
In security work that absence is the point. Probing a system for holes looks identical whether an intruder is doing it or the specialist hired to check the locks, so a filter that stops the first one stops the second one too.
Nobody gets Mythos into their own hands. It goes to companies whose business is the security of other people’s systems. Alongside it came the Defender Advantage fund: 35 million dollars in Claude credits for the people patching holes in open source.
The caution has a history. In late July Anthropic described incidents where Claude models slipped out of the lab during security testing and went after real systems. A model that can find a vulnerability and write the patch can just as well find one and leave it alone – and this one has nothing left inside it to stop the second thing.
Spellchecking on the horizon!
Version 2.1.235 began rolling out spellcheck, which underlines misspelled words in the box where you type to Claude Code. It has half arrived: there is no switch in /config, the settings documentation has no entry for the key, and it needs a dictionary installed separately on your system. So if you go looking and nothing gets underlined, nothing is broken – it simply has not reached you yet. We will come back to it once it shows up in the settings panel.
Claude Academy: the courses date from March, the address is new
Anthropic spent last week showing off its own course platform. The company blog post from 20 August sets out the principles behind how it designs its material and promises broad access to it.
The catch is that this is not a catalogue of new courses – those have been available since March 2026. What changed is the layout of the site and its address: the courses now sit at academy.claude.com, while the old anthropic.skilljar.com still answers and still lists the lot. Which means that if you know these courses already, the Academy has little new to show you.
The agent-building tools leave beta
This section is for people building their own applications on the Claude model – there is nothing to switch on in Claude Code. Computer use, meaning control of a computer screen, along with the Skills API and the Files API, stopped being beta releases. Browser use arrived separately: an extension of computer use that adds the structure of the page to the picture of the screen, so an agent lands on an actual form field instead of aiming at pixels.
Asteroid Labs, which tested computer use, reports that its longest claims workflow came down from 32 minutes to 13, and that the cost of a single task fell by roughly 30 percent. When some service promises that its AI will fill in a form for you, this is the machinery underneath.