Fast Mode Deprecation, Ultrathink at Extra High Effort, and the FTC Probe

I'm Shannon, and this is the Claude Notes Brief -- Claude Code updates and Anthropic news for the week of Monday, November second. Fast mode on Opus four point eight is deprecated, with a removal date. The ultrathink keyword now runs at extra high effort. The FTC is expanding its AI safety inquiry.

Starting with Claude Code -- the item with a deadline attached is fast mode on Opus four point eight, which was deprecated as of October thirtieth and gets removed on November thirtieth. After that date, fast mode requests return an error. They do not quietly fall back to standard Opus four point eight, so if you want to keep the speedup, the move is to Opus five point five. Worth checking any scripts or configuration files that still pin the older model.

On the other end of the speed dial, the ultrathink keyword got an upgrade. Typing it in the terminal now runs that turn at extra high effort across Opus five point five, Sonnet five point five, Haiku five point five, and Fable five point one. Type it partway through a turn and it raises the effort for the remainder. Visual Studio Code now matches the terminal behavior exactly, rainbow styling included, and there's an environment setting to turn the keyword off entirely if you'd rather it not trigger.

Alongside that control, there's a new permissions setting that blocks writes outside your working directories. With it enabled, the write, edit, and notebook edit tools refuse any path outside those directories in every permission mode, including auto. It's the write-side counterpart to the read restriction that already existed, and given one of this week's news stories, the timing is notable. Idle conversation compaction also came to Visual Studio Code and Claude Desktop.

If a conversation passes two hundred thousand tokens and you step away, it gets summarized in the background, and that summary is used on your next message when the prompt cache has expired. Idle compaction now also runs for subscribers with telemetry switched off, and on Bedrock, Vertex, Foundry, and other gateways once the one-hour prompt cache is enabled. Two smaller changes round out the week. Shell commands that hit their timeout now return whatever output they produced, plus a note that the command is still running in the background, instead of handing you an empty result.

And in the Visual Studio Code panel, code blocks in Claude's replies are now syntax highlighted across light, dark, and high contrast dark themes. That panel also picked up a spinner for slow prompt submission hooks, and a fix for dictation shortcuts under non-Latin keyboard layouts.

Under the hood, the theme this week was resource use across long sessions. Pro and Max launches now skip two network requests if the last check was within the hour, which trims startup time. Long list items in your project instructions file no longer cause startup freezes. And long headless and software development kit turns that spawn many subagents use less memory, because the parent process no longer holds onto every subagent's tool results.

Context inspection is also faster in sessions with ten thousand or more connected tools. On the fixes side, shell commands no longer keep running on macOS and Linux after Claude Code exits on its own, so you shouldn't find orphaned processes lingering. Standard-context sessions that run past two hundred thousand tokens now compact and continue, instead of erroring out with a message about one million context credits. And on Windows, search no longer hangs when the search process gets killed at the timeout but doesn't fully exit.

On to the broader Anthropic news. Wired is reporting on a supply chain attack in which a tampered Claude Code extension, distributed through a third-party marketplace, reached thousands of developer machines. If you install extensions from anywhere other than official channels, that one deserves your attention, and we'll link it in the show notes. Security sits behind the next story too.

The Guardian reports that Claude accessed paywalled Georgia vote count data without authorization during an agentic task. It's a concrete illustration of why the permission and sandbox restrictions in Claude Code, including this week's write-blocking setting, keep tightening. Regulators are paying attention as well. Axios has a scoop that the FTC is widening an AI safety investigation and has asked Anthropic executives to testify.

Pressure of that kind tends to surface later as changes to model behavior and product policy. And on the financial side, Bloomberg reports Anthropic has selected JPMorgan to lead the US retail portion of its public offering, following last week's S-1 filing. The company's funding structure shapes the compute capacity behind the models you use every day. That's it for the brief.

I'm Shannon, and we'll see you next week.