Everything Claude Shipped
in March 2026
Code Review agents on every PR, a 1M context window for free, native Computer Use on Mac, Auto Mode that handles its own permission loops — March was a big month. Here's every feature that shipped, what it means, and how to use it today.
TL;DR — March 2026 in 60 Seconds
- ▸ Code Review: open a PR → Claude sends a team of agents to find bugs before you even look
- ▸ Auto Mode: Claude handles its own permission prompts safely — no more y/n loops
- ▸ Computer Use: Claude opens files, navigates browsers, runs dev tools directly on your Mac
- ▸ 1M context window: now the default for Opus 4.6 on paid plans, no surcharge
- ▸ Mythos (Capybara): next-gen model leaked, confirmed real — step-change in coding + reasoning
Claude Code Review New
Launched ~March 21, 2026
When a pull request opens, Claude Code dispatches a team of specialized agents to review it — one hunts for logic errors, one checks edge cases, one scans for security issues. The review appears as a comment thread on the PR, not a wall of text. You get findings grouped by severity, with suggested fixes inline.
How to enable it
Add the claude review command to your CI workflow, or install the GitHub App from your Claude settings dashboard:
# .github/workflows/claude-review.yml
name: Claude Code Review
on: [pull_request]
jobs:
review:
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- name: Run Claude Code Review
run: claude review --pr ${{ github.event.number }}The multi-agent design matters: each agent gets the full diff but focuses on a distinct concern, then results are merged and deduplicated. In practice this means fewer false positives than a single-pass LLM review and much faster turnaround — typically under 90 seconds for diffs under 500 lines.
What it catches well
- ▸ Off-by-one errors and missed null checks
- ▸ SQL injection patterns and unsanitized inputs
- ▸ Inconsistencies between new code and existing test coverage
- ▸ API contract violations (wrong field names, wrong HTTP methods)
Auto Mode Research Preview
Launched March 24, 2026
The biggest friction in Claude Code's agentic mode was the permission loop: every file write, every shell command triggered a y/n prompt. Auto Mode replaces this with a safety classifier that evaluates each action's risk level before deciding whether to proceed, pause, or escalate.
How the classifier works
# Enable Auto Mode in claude settings
claude config set autoMode true
# Or per-session
claude --auto "refactor the auth module to use JWT"It's marked as a research preview, meaning Anthropic is actively collecting failure reports. If Auto Mode makes a surprising decision (auto-proceeds when you expected a pause, or blocks something safe), you can flag it inline with /feedback. The classifier improves from these signals.
Computer Use on Mac New
Launched March 23, 2026 · Pro & Max plans
Claude Code and Claude Cowork can now directly interact with your Mac — open files in Finder, navigate browser tabs, fill forms, click buttons in dev tools, and run terminal commands — without any additional driver setup. The model sees your screen via a screenshot loop, plans the next action, executes it, then re-screenshots.
Enabling it
- Open Claude Code → Settings → Computer Use
- Grant Accessibility and Screen Recording permissions (macOS System Settings)
- In any Claude session: /computer-use on
Good use cases
- ▸ Browser testing without Playwright setup
- ▸ Navigating legacy GUI-only internal tools
- ▸ Automating repetitive dashboard clicks
- ▸ Debugging in Chrome DevTools visually
Current limitations
- ▸ macOS only (Windows/Linux: roadmap)
- ▸ Slower than headless Playwright for pure automation
- ▸ Requires Pro or Max plan
- ▸ Not available in shared/multi-user machines
1M Token Context Window — Now Default GA
Generally available March 13, 2026 · No surcharge
The 1 million token context window for Claude Opus 4.6 and Sonnet 4.6 is now generally available at standard pricing — no extended-context surcharge. For Max, Team, and Enterprise plans it's the default (Opus 4.6 uses 1M by default via Claude Code 2.1.75). Media limits also jumped from 100 to 600 images per request.
What 1M tokens actually means for developers
In practice: you can now drop an entire mid-size monorepo into context and ask Claude to trace a data flow end-to-end, find all callers of a deprecated method across hundreds of files, or review migration compatibility across a full codebase — all in one session without any chunking strategy.
# Load an entire repo into context
claude --include "**/*.py" --include "**/*.ts" \
"Find every place we call legacy_auth() and map out all the callers"
# Or using project files
claude --project . "Trace the full data flow from POST /api/orders to the database"Claude Code 2.1.75 Changelog Release
~March 22, 2026
Version 2.1.75 shipped 19 CLI changes and 2 system prompt changes. Here are the ones that matter day-to-day:
1M context window default on paid plans
Opus 4.6 now defaults to 1M tokens on Max, Team, and Enterprise. Previously required explicit --max-tokens flag.
Smarter permission denial messages
When you deny a tool use and your intent is unclear, Claude now asks "what were you trying to avoid?" instead of making a guess and doing the wrong thing.
Faster tool-call streaming
Tool results (bash output, file reads) now stream inline instead of buffering. For long-running commands you see partial output immediately.
/compact threshold lowered
Context compaction now triggers at 85% window usage instead of 95%, reducing mid-task interruptions on long sessions.
Interactive Charts & Diagrams Beta
All plans including free
Claude can now build interactive charts and diagrams directly inside the chat window. You can hover for details, click to filter, and resize. It works on claude.ai and in the Claude iOS/Android apps. No external tool or export needed.
Supported output types (beta)
For developers this is most useful during data exploration — paste a JSON payload or CSV, ask Claude to visualize it, iterate on the chart in-place. The rendered artifact is exportable as SVG or PNG via the three-dot menu.
Claude Code Channels (Telegram + Discord) New
Launched ~March 20-21, 2026
Claude Code Channels lets you hand off tasks to your coding agent via Telegram or Discord — from your phone, while away from your desk. The agent runs async on your machine (or a remote container) and pings you back when done or when it needs a decision.
Setup in 3 steps
# 1. Install the Channels MCP server
claude mcp install channels
# 2. Connect your Telegram or Discord account
claude channels connect telegram
# 3. Start a session from mobile — Claude runs on your machine
# Message @ClaudeCodeBot on Telegram:
# "Fix the failing tests in /auth and open a PR"Code with Claude Conference — 2026 Upcoming
Anthropic is running their developer conference across three cities this spring. Full-day format: workshops, live demos, and 1:1 office hours with the teams behind Claude Code, the API, and MCP.
The livestream is free for all plans. In-person registration: claude.com/code-with-claude →
Claude Mythos — The Leaked Next-Gen Model Breaking
Leaked March 26, confirmed March 27, 2026
An internal CMS misconfiguration exposed ~3,000 unpublished assets including a draft blog post about a new model internally called Claude Mythos, tier name Capybara. Anthropic confirmed it's real within hours of the leak spreading.
What Anthropic confirmed
- ▸ Training is complete — not a future roadmap item
- ▸ Described internally as a "step change" vs all prior models
- ▸ Dramatically higher scores on coding, reasoning, and cybersecurity benchmarks
- ▸ Anthropic flagged it as posing "unprecedented cybersecurity risks" — hence a controlled rollout
- ▸ Currently in limited early access with select customers — no public API date yet
The "cybersecurity risk" framing isn't alarming in context — Anthropic has consistently applied extra scrutiny before releasing models with elevated offensive security capabilities. The controlled rollout likely means a red-teaming period before broader access. Given the March training completion, a public beta in Q2 2026 is plausible.
Sources: Fortune · The Decoder · SiliconAngle
Also Shipped This Month
Doubled off-peak usage limits
Free, Pro, Max, and Team plans got 2x usage outside peak hours for two weeks. Weekend and post-2pm sessions got 10-hour limits.
Anthropic Science Blog launched
New blog documenting AI-driven research — first post: physicist Matthew Schwartz on supervising Claude through theoretical physics calculations.
The Anthropic Institute
New research arm focused on AI's societal challenges, led by co-founder Jack Clark. Focus on economic impacts and alignment research.
Anthropic Bengaluru office
India office opening in early 2026. First presence in South Asia — expected to focus on developer ecosystem partnerships and research hiring.
DoD injunction (First Amendment)
Federal judge granted Anthropic a preliminary injunction blocking the Pentagon's blacklisting, citing First Amendment retaliation. CEO Dario Amodei published "The Adolescence of Technology" essay on AI national security risks.
IPO consideration
Bloomberg reported Anthropic is weighing an IPO as early as October 2026, with projected 2026 revenue of $30B. No formal announcement yet.
Feature Priority Matrix
Which features to enable today vs. watch:
| Feature | Plans | Action |
|---|---|---|
| 1M Context Window | All paid | Default on Max/Team |
| Code Review | All paid | Enable in GitHub settings |
| Interactive Charts | All plans | Auto — just ask |
| Auto Mode | All paid | Try cautiously |
| Code Channels | All paid | Install MCP server |
| Computer Use | Pro & Max | Enable + grant permissions |
| Claude Mythos | Early access | Waitlist only — watch for GA |
Get the next roundup in your inbox
Developer-focused AI updates — no hype, just what shipped and how to use it.