Sonnet 5.5 Becomes the Default, and Claude Mods Arrive

I'm Shannon, and this is the Claude Notes Brief -- your weekly rundown of Claude Code updates and Anthropic news for the week of Monday, October fifth. Claude Sonnet five point five is now the default Sonnet model. Plugins can rewrite prompts and block risky commands. And Anthropic is reportedly targeting a listing before Thanksgiving.

Let's start with Claude Code, where Claude Sonnet five point five is now the default Sonnet model on the Anthropic API, and it comes with a one million token context window. If you run Sonnet for subagents or everyday coding, you get that larger context automatically -- there's no special long-context variant to pick anymore, and the Sonnet alias resolves to the new model on its own. Pricing sits at two dollars per million input tokens and ten dollars per million output, with cache reads at twenty cents per million. That default-to-a-million theme carries over to enterprise deployments too.

Opus four point seven and later, plus Fable, now default to one million tokens of context on Bedrock, Vertex, Foundry, the Claude apps gateway, and behind a custom base URL. If your gateway caps out at two hundred thousand, there's an environment setting and an autocompact option to pull it back down. The other headline this week is Claude Mods, which open up a much bigger customization surface. Mods are TypeScript functions that hook directly into Claude Code itself -- you can rewrite prompts before they're sent, block commands you consider risky, draw your own custom panes, or replace built-in features entirely, then ship the whole thing as a plugin.

There's a built-in example called You Should Know that runs a side agent flagging things you or Claude might have missed, and you can switch it on from the plugin menu. Anthropic published a companion writeup on the Claude blog walking through how to build one, and we'll link it in the show notes. Permissions got a default change as well. Interactive terminal and Visual Studio Code sessions now start in auto mode when you haven't configured a permission mode yourself -- that applies on every plan and every provider, and headless and Python SDK sessions follow the same rule.

If you've set a default mode explicitly, that still wins. The Visual Studio Code extension picked up bookmarks, so you can save Claude's responses to a side panel, and you can now move a running command or subagent into the background right from the chat. Background shell and Monitor output shows up on the cards in the agent map. And stopping now ends only the current turn, leaving background agents running until you stop each one.

A few smaller additions round things out. If you wipe a draft prompt, pressing up on the empty prompt brings it back, including pasted text and images. There's a new find-by-name shortcut in the agents view, keys for jumping between groups, a flag that opens the Claude desktop app on your current directory, and a new command for viewing or saving a plugin's configuration options.

Under the hood, this week was about closing permission gaps and trimming overhead. On the security side, a dangerous remove command tucked inside a bash or shell script no longer runs unprompted under bypass permissions or a shell allow rule. Deny and ask rules now catch commands hidden behind environment variable prefixes under sandbox auto-allow. And logs and transcripts stop leaking fragments of URL passwords and keys that were obscured by invisible characters.

On the performance side, Bash commands lost a subshell that was running before every single invocation, so those calls are faster now. Startup only builds the settings schema your configuration actually uses. Large results coming back from MCP servers use less memory and produce smaller session files. And that brief freeze some of you hit partway through long sessions outside fullscreen mode is gone.

Now for the broader Anthropic news, and most of it points in one direction. Bloomberg is reporting that Anthropic is targeting a public listing before Thanksgiving. Reuters covered the Sonnet five point five launch through that same lens, framing the pricing strategy against the company's path to a listing. For developers, a public Anthropic means a lot more disclosure about the economics behind the models you run every day.

Running alongside that, the New York Times reports the Federal Trade Commission has opened an inquiry into both OpenAI and Anthropic over potential consumer harms. Regulatory attention like this tends to surface later as tighter default permissions and more safety classifiers in agentic tools -- which makes this week's auto mode default and permission fixes worth noting in context. And CNBC reports Anthropic has committed one hundred million dollars to train ten thousand engineers through something called Claude Frontier Academy, running through the end of twenty twenty-seven. It's aimed at the enterprise deployment skills gap.

On a related note, Claude for Government reached general availability this week, with the Claude Code command line interface in early access -- relevant if you work in federal environments where procurement has been the blocker. That's it for the brief. I'm Shannon, and we'll see you next week.