Issue #3

February 24, 20265 min read

Claude Sonnet 4.6 launches with 1M context window

Claude Sonnet 4.6 launches with a 1M context window, alongside git worktree isolation via claude -w and real-time VS Code plan previews. This week covers Bloomberg's piece on Claude Code's impact on Anthropic, the Sonnet 4.6 release, a Pentagon AI safety dispute, and market reaction to Claude Code Security.

The Surprise Hit That Made Anthropic Into an AI Juggernaut — Bloomberg

On this page

Highlight
Worth Knowing
Under the Hood
From Anthropic
In the News

Highlight

Claude Sonnet 4.6 Arrives With Full 1M Context Window

Claude Sonnet 4.6 is now available and replaces Sonnet 4.5 on the Max plan -- if you're on Max, switch via /model to get the new model. Sonnet 4.6 brings the full 1M context window to the mid-tier model, and fast mode on Opus 4.6 now also gets 1M context. If you want to opt out of 1M context window support for any reason, you can set the new CLAUDE_CODE_DISABLE_1M_CONTEXT environment variable.

Worth Knowing

Git Worktree Isolation Now Available for Claude and Subagents

You can now launch Claude in an isolated git worktree with claude -w, keeping experimental work separate from your main branch. Individual agent definitions can also declare isolation: "worktree" to always run in their own worktree, and the new WorktreeCreate and WorktreeRemove hook events let you run custom setup and teardown scripts around that lifecycle.

VS Code Plan Preview Now Updates in Real Time

If you use VS Code, the plan preview now refreshes live as Claude works through iterations, rather than requiring a manual refresh. Commenting is locked until the plan is ready for review, and if you reject a plan, the preview stays open so Claude can revise it in place.

Simple Mode Gets File Editing and a Fully Stripped Configuration

When you set CLAUDE_CODE_SIMPLE, Claude now strips out MCP tools, attachments, hooks, and CLAUDE.md loading entirely -- giving you a predictable, minimal environment. At the same time, the file edit tool is now included alongside Bash, so you can still make direct file edits in simple mode.

New CLI and Hook Additions for Agent and Config Management

You can now run claude agents to see all configured agents at a glance. If you write hooks, the new ConfigChange event fires whenever config files change mid-session -- useful for auditing or enforcing policy -- and Stop and SubagentStop hooks now receive a last_assistant_message field so you can read the final response without parsing transcript files.

Background Agent Controls Reorganized Around Ctrl+F

Pressing Esc no longer kills background agents -- they now keep running when you cancel the main thread. To kill all background agents, press Ctrl+F twice within 3 seconds. Agent definitions also now support a background: true field to always run that agent as a background task.

Under the Hood

Extensive Memory and Startup Performance Work This Week

This release cycle includes a large batch of memory leak fixes -- covering agent task state, LSP diagnostics, shell commands, Yoga WASM, and circular buffers -- alongside startup improvements that reduce time-to-interactive by around 500ms in headless mode (-p) and cut redundant network calls during MCP initialization.

Session Continuity, Permissions, and Windows Bugs Fixed

Several reliability fixes landed this week: plan mode and custom session names now survive context compaction, sessions with large first prompts (over 16KB) no longer disappear from /resume, the bash permission classifier can no longer hallucinate rule descriptions to grant itself permissions, and hooks now execute correctly on Windows via Git Bash.

From Anthropic

Claude Blog · 2 min read

Bringing automated preview, review, and merge to Claude Code on desktop

Anthropic's post walks through the new desktop workflow for previewing running apps, auto-reviewing diffs, and monitoring PRs -- all without leaving Claude Code. If you've been context-switching between your editor and GitHub, this describes what the integrated flow now looks like.

Claude Blog · 2 min read

Bringing automated preview, review, and merge to Claude Code on desktop

Claude Blog · 3 min read

Increase web search accuracy and efficiency with dynamic filtering

Claude's web search and fetch tools now use code execution to filter results before they hit your context window -- Anthropic reports an average 11% accuracy improvement and 24% fewer input tokens. If you use web search in your Claude Code workflows, this should reduce noise and cost.

Claude Blog · 3 min read

Increase web search accuracy and efficiency with dynamic filtering

In the News

Bloomberg · 6 min read

The Surprise Hit That Made Anthropic Into an AI Juggernaut

Bloomberg traces how Claude Code -- released roughly a year ago -- became the product that defined Anthropic's trajectory and pushed rivals to accelerate their own coding tools. Worth reading for context on how the tool you use daily became central to one of the most-watched AI companies.

Bloomberg · 6 min read

The Surprise Hit That Made Anthropic Into an AI Juggernaut

Bloomberg · 4 min read

Anthropic says new AI model is better at using computers

Bloomberg's coverage of the Sonnet 4.6 launch focuses on its improved ability to handle multi-step computer tasks -- the same model now available in Claude Code as this week's highlight.

Bloomberg · 4 min read

Anthropic says new AI model is better at using computers

Bloomberg's coverage of the Sonnet 4.6 launch focuses on its improved ability to handle multi-step computer tasks -- the same model now available in Claude Code as this week's highlight.

New York Times · 5 min read

Defense Department and Anthropic Square Off in Dispute Over A.I. Safety

The NYT reports on the escalating dispute between Anthropic and the Pentagon over use restrictions on Claude -- a conflict that touches directly on the usage policies governing the models in Claude Code.

New York Times · 5 min read

Defense Department and Anthropic Square Off in Dispute Over A.I. Safety

Bloomberg · 4 min read

Cyber Stocks Slide as Anthropic Unveils 'Claude Code Security'

Bloomberg covers the market reaction to Anthropic's new Claude Code security feature, which spooked cybersecurity stocks on Friday -- a signal of how seriously the industry is taking Claude Code's expanding capabilities.

Bloomberg · 4 min read

Cyber Stocks Slide as Anthropic Unveils 'Claude Code Security'

With Sonnet 4.6 now the default on Max and a full 1M context window across both Sonnet and Opus fast mode, this is a good week to check your /model setting and make sure you're on the current model.