Opus 4.6 Gets One Million Tokens by Default, MCP Servers Learn to Ask Questions

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 sixteenth. Opus four point six defaults to one million tokens of context. MCP servers can now pause and ask you questions. And Wired goes inside OpenAI's race to catch Claude Code.

Five releases for Claude Code this week, headlined by a major context window expansion -- Opus four point six now defaults to the full one million token context window on Max, Team, and Enterprise plans. No extra usage tier, no toggle. If you're on one of those plans, it's already live. That means longer sessions, larger codebases, and fewer compaction interruptions before you hit the ceiling.

This ties into Anthropic's broader announcement that one million token context is now generally available for both Opus four point six and Sonnet four point six, and the long-context pricing premium is gone entirely. Standard pricing applies across the full window. That's a meaningful cost change if you regularly work with large projects. Now, a bigger context window is only useful if you can see what's filling it up, and that's exactly what a new diagnostics feature addresses.

Running the context command now analyzes your session and flags specific issues -- bloated memory files, context-heavy tools, capacity warnings -- along with actionable tips on how to free up space. If you've been guessing why compaction kicks in earlier than expected, this gives you concrete answers. The other headline feature this week is MCP elicitation. MCP servers can now pause execution mid-task and prompt you for structured input -- a form with fields, or a browser URL for an OAuth flow -- through an interactive dialog.

If you build or use custom MCP integrations, this unlocks multi-step workflows that previously required clunky workarounds. New hooks let you intercept and override those responses programmatically too. Rounding out the week, the plan command got a quality-of-life upgrade. You can now pass a description directly -- something like "fix the auth bug" -- and it enters plan mode and starts working immediately, saving you the extra prompt.

There's also a new effort command that lets you adjust model effort level without leaving your session. And for teams working in large monorepos with worktree mode, a new sparse paths setting lets you specify exactly which directories to check out using git sparse checkout, so you're not pulling the entire repo into every worktree.

Under the hood, this was a week of meaningful stability and performance work. A memory leak in streaming API response buffers that caused unbounded memory growth on the npm code path is now fixed. Token estimation for thinking and tool use blocks was over-counting, which triggered premature compaction -- that's corrected. And SDK query calls now maintain prompt cache correctly, which Anthropic says reduces input token costs by up to twelve x in some cases.

That cost reduction pairs nicely with the removal of the long-context pricing premium. On the stability side, permission prompts for complex bash commands were triggering full CPU loops or freezes -- that's resolved. The team also fixed a deadlock that occurred when many skill files changed at once, stopped plan mode from requesting redundant re-approval, and made auto-compaction stop after three consecutive failures instead of retrying indefinitely. Separately, several common read-only commands like lsof and pgrep are now on the bash auto-approval list, so you'll see fewer permission prompts during normal work.

Stepping outside Claude Code, there's a lot moving in the broader Anthropic world this week. Wired published a detailed profile of OpenAI's efforts to catch up with Claude Code through its Codex product. The piece offers a rare look at how the competition views the agentic coding landscape, and it's a good read if you're curious about where these tools are headed. We'll link it in the show notes.

On the enterprise side, Reuters is reporting that Microsoft is integrating Anthropic's models into its Copilot Cowork product, its push into AI agents. That's a significant distribution channel -- it puts Claude's underlying models in front of a much wider enterprise audience beyond direct API usage. And TechCrunch reports that Anthropic has launched a dedicated code review tool aimed specifically at verifying AI-generated code. As more code gets written by agents, a purpose-built review layer feels like a natural complement.

Finally, the legal situation between Anthropic and the Pentagon continues to develop. According to Reuters, Anthropic is now seeking an appeals court stay of the supply-chain risk designation, which could affect its ability to serve government-adjacent enterprise customers. That story is still unfolding, and we'll keep tracking it. That's it for the brief.

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