Opus 4.6 Goes Default, /loop Lands, and the Pentagon Fallout Continues

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of March ninth. Opus four point six becomes the default model. A new loop command brings recurring prompts to your sessions. And the Pentagon fallout gets a financial counterpoint from Bloomberg.

Five releases for Claude Code this week, headlined by Opus four point six becoming the default model for Max and Team plan users. This is a medium effort setting -- a balance between speed and thoroughness that Anthropic says works well for most tasks. If you had Opus four or four point one pinned, you've been automatically migrated to four point six. And if you want to push the model harder on a specific turn, the ultrathink keyword is back for forcing high effort.

Sonnet users saw a similar migration -- anyone on Sonnet four point five has been moved to Sonnet four point six automatically. That model change is the headline, but the workflow improvements around it matter just as much. There's a new loop command that lets you schedule recurring prompts inside a session. You tell Claude what to do and how often -- say, every five minutes, check the deploy -- and it'll keep running that prompt until you stop it.

Cron-style scheduling is also available for more complex patterns. It pairs well with a new blog post from the Claude team on common workflow patterns for AI agents, which we'll link in the show notes. Moving over to the editor side, Visual Studio Code picked up a meaningful set of additions. There's now a spark icon in the activity bar that lists all your Claude Code sessions, and each one opens as a full editor tab.

Plans render as full markdown documents where you can add comments to steer Claude's approach. And you can manage MCP servers directly from the chat panel -- enabling, disabling, reconnecting, and handling OAuth without switching to the terminal. Voice mode also got a significant expansion this week, adding ten new languages -- Russian, Polish, Turkish, Dutch, Ukrainian, Greek, Czech, Danish, Swedish, and Norwegian -- bringing the total to twenty. You can also rebind the push-to-talk key now, and modifier combos help avoid interfering with normal typing.

Rounding out the feature side, there are two new settings worth knowing about. One lets you strip Claude's built-in git instructions from the system prompt, which frees up context if you have your own commit and PR conventions. The other relaxes sandbox networking on macOS for corporate environments where a man-in-the-middle proxy breaks TLS verification for tools like the GitHub CLI or Terraform.

Under the hood, this was a big week for stability and memory. Baseline memory usage drops about sixteen megabytes thanks to deferred WebAssembly loading, and prompt input re-renders are down roughly seventy-four percent. Several memory leaks in long sessions have been fixed too -- including a React render scope issue that accumulated around thirty-five megabytes over a thousand turns, and an in-process teammate history bug that prevented garbage collection after clearing a session. Those long-running sessions should also no longer freeze on keystroke input.

A startup hang caused by voice mode's CoreAudio initialization -- five to eight seconds in some cases -- is resolved, along with a separate hang from simultaneous OAuth refreshes. If you use SSH, pressing Enter no longer inserts a stray newline over slow connections. Windows and WSL clipboard handling now works correctly with CJK characters and emoji. And API four hundred errors that cropped up with third-party gateways and custom Bedrock inference profiles have been fixed.

Turning to the broader Anthropic news, the Pentagon story continues to develop. Reuters published a detailed reconstruction of Anthropic's Pentagon negotiations -- tracing the timeline from initial courtship through the breakdown over AI safeguards. That's useful context if you've been following the dispute since last week's formal blacklist. The practical question for most developers, though, is whether their access is affected.

CNBC reports that Amazon, Microsoft, and Google have all confirmed Claude remains available to non-defense customers on their cloud platforms. If you access Claude through AWS, Azure, or Google Cloud, nothing changes for you. On the financial side, Bloomberg reports that Anthropic is approaching a twenty billion dollar annualized revenue run rate -- even amid the Pentagon conflict. That's a signal the company's commercial business continues to grow independently of the defense sector.

And finally, the Wall Street Journal has a story on Claude finding twenty-two security vulnerabilities in Firefox's codebase. It's a concrete demonstration of agentic coding applied to a large real-world project, and it's directly relevant if you're thinking about using Claude Code for security review work. Anthropic's engineering team also published a companion piece this week on how infrastructure configuration alone can swing agentic coding benchmarks by several percentage points -- sometimes more than the gap between top models on the leaderboard. Both are well worth your time, and we'll link them in the show notes.

That's it for the brief. I'm Shannon, and we'll see you next week.