One of this week’s changes may already have caught your eye: in auto mode, Claude Code started asking whether it may read files that sit outside the folder you work in. It is worth knowing which answer to give, because two of the three stay with you for good.
Anthropic shipped Claude Fable 5.1
Claude Fable 5.1 came out on Tuesday 1 September and is the strongest model you can pick in Claude Code today. You select it with the /model command, exactly as you did its predecessor. You need Claude Code 2.1.255 or newer.
In Fable 5.1 Anthropic improved what the model can do and, more to the point, its safety filters. Fable 5 was over-cautious and would refuse to answer basic questions from biology or medicine. Fable 5.1 does that markedly less often. Anthropic reports that its cyber safeguards interrupt work around 60 percent less often per session than in Fable 5, and that the new biology and medical filters – which went into Fable 5 as well – fire 85 percent less often on harmless questions.
Using Fable got cheaper too. For re-reading the context a conversation has already stored, Anthropic charges separately with every follow-up answer – that is a cache read. In Fable 5.1 the rate dropped from a dollar to 25 cents per million tokens, three quarters off, and the rest of the Fable price list stayed where it was. Anthropic puts that at around 25 percent off the cost of ordinary work, and up to 45 percent on tasks where the AI works for long stretches with nobody watching.
Two days after Fable 5.1, on Thursday 3 September, OpenAI released a new model of its own. The company calls GPT-6 Astra the most intelligent model in the world. We are not going to settle which of the two is better – what we do like is that the fight between two giants carries on, because it keeps producing better models.
Back to Anthropic: on the Pro plan both Fable models still cost extra, on top of the subscription. On Max they are included, but they can take at most half of your weekly limit – the same limit every other model draws from.
Which model suits which kind of task is the subject of our tip Haiku, Sonnet, Opus, Fable – which model for what. And if you are still choosing a plan, have a look at Pro or Max – which Claude plan to start with.
In auto mode, Claude Code asks before it leaves the working folder
Since version 2.1.257, released on Tuesday 1 September, Claude Code asks for permission the first time it reads a file outside the folders it works in. This concerns auto mode, the one an ordinary terminal session starts in. Until now the AI reached out there without a word, because for reading, the working folder was no boundary at all.
The question has three answers: Keep allowing, Block from now on and Ask again next time. The first lets this read through and every later one, and Claude Code records the answer, so the question does not come back. The last one puts the decision off: it refuses this read, but asks again at the next one. The answer with the longest reach is Block from now on. That answer refuses the read too, but on top of it writes a ban into your settings – a ban on every read outside the working folders, not just on this one file. The ban then holds in every later session and in every permission mode, not only in auto mode.
If your home folder, next to the one you work in, holds contract templates, price lists and notes from meetings, we recommend Block from now on. It is an ordinary setting, so you can lift it later, and a folder you need for one particular task you add with the /add-dir command.
/skill-doctor, a new way to examine your skills
Claude Code has a new command, /skill-doctor, which shows how much context each of your skills takes up and how often it fires. The report points out separately the ones Claude Code never reached for, and tells you where they are switched off. It also lists plugins nobody has touched in a long time. The same report opens in the /plugin manager, in the Stats tab. The command works from version 2.1.252.
Every skill on the list adds itself to the context on every turn of the conversation, whether or not Claude reaches for it. A dozen skills kept just in case therefore take up room in conversations about something else entirely.
If your first skill is still ahead of you, our lesson Your first skill – a morning plan for the day goes through ordering it from Claude Code step by step, and what to fix in it afterwards.
More ways to handle MCP in VS Code
The MCP servers dialog in the Claude Code extension for VS Code gained two new actions: Add and Remove. The change came in version 2.1.261, released on Friday 4 September. You add or remove an MCP server straight from the list in that dialog, and the dialog itself opens with the /mcp command, typed into the conversation panel. Extension interfaces change often, so yours may already look a little different.
For anyone working in VS Code, both actions are mainly a convenience: everything to do with MCP servers now happens in one dialog. Anyone working in the terminal alone will not find that dialog – there you still add an MCP server by asking for it in the conversation. Asking works the same everywhere, as it happens, so if you work in both the terminal and VS Code, you can stay with it.
What an MCP server even is, and how it differs from a connector on claude.ai, is the subject of our tip Connector or MCP: what actually makes them different.
The time format is yours to set in /config
Claude Code shows the time in two places: in the message that closes each answer, and in the timestamps in the transcript view. From version 2.1.257, released on Tuesday 1 September, you can set the format of that time to your own taste. The /config panel gained a Time format row, with four options on it: auto, 12-hour, 24-hour and 24-hour-utc.
It is easy to confuse 24-hour with 24-hour-utc, and the difference is a real one. The 24-hour option shows local time, the time from your computer’s clock or from a time zone set separately. The 24-hour-utc option converts it to universal time and sticks a Z after the minutes, so 18:05 on a Central European clock in summer shows up there as 16:05Z.
The auto option leaves everything as it was: the time follows your system’s regional settings, so if those already write a 24-hour clock, the auto option changes nothing for you. The 24-hour-utc option earns its place when you line up times with somebody abroad, or with a log in which everything is in universal time. Your own pattern and a time zone are not on the panel’s list – both go into the settings file, and for that it is enough to ask:
Set my time format in Claude Code to 24-hour and my time zone to Europe/London.What else in the /config panel is worth changing right away, and what to leave alone, is in our tip Claude Code settings – what to change, what to leave.
To close: the complete map of a fruit fly’s brain
On Thursday 3 September the HHMI Janelia team, together with Google, published the first complete map of the nerve connections in a male fruit fly: 166,000 neurons and 125 million synapses. It is the largest brain map by number of neurons anyone has put together so far. It covers the insect’s entire central nervous system: the brain, both optic lobes and the nerve cord.
The work that could not have been finished by hand was done by neural networks. From electron microscope images they reconstructed the shape of every cell and the points where it joins its neighbours, and people checked the result. Google Research writes plainly that only cutting down the manual error correction brings bigger brains within reach. The finished dataset is free and opens in a browser, in Neuroglancer, with no account and nothing to install.
A human brain has 86 billion neurons, so the researchers note straight away that this is still a very long way off. How many projects like it are standing half-finished today, waiting not for an idea but for hands to do the correcting?