Claude Code launched a wave of updates throughout March 2026 — 74 feature releases in 52 days, jumping from version 2.1.63 to 2.1.76. If you haven't kept up, you're missing tools that fundamentally change how you write, monitor, and ship code. This guide covers every major addition and tells you exactly how to use each one.

What Changed in Claude Code v2.1.63–v2.1.76 (March 2026)

Here's the quick-reference breakdown before we go deep:

---

1. The /loop Command: Turn Claude Code Into a Background Worker

Released: March 2026 (v2.1.71)

The /loop command is the biggest quality-of-life upgrade in this batch. It transforms Claude Code from a reactive coding assistant into a continuously running background agent that checks things for you on a schedule.

Basic Syntax

/loop [interval] [natural-language task]

You define how often (5m, 30s, 1h) and what to do. Claude handles the rest.

Real-World Examples

Monitor a deployment:

/loop 5m check if the staging deploy succeeded and report any errors

Watch open pull requests:

/loop 15m check all open PRs for new review comments or CI failures

Run your test suite on a timer:

/loop 30s run the unit tests in /src/tests and summarize failures

Daily code quality audit:

/loop 24h scan the src/ directory for security anti-patterns and log results

Limitations to Know

• Maximum 50 concurrent loops per session
• Loops expire automatically after 3 days
• All loops stop when the session closes
• Can be disabled globally with CLAUDECODEDISABLE_CRON=1

Why It Matters

Before /loop, monitoring CI/CD pipelines or watching deployments required you to manually context-switch from coding, open a browser or terminal tab, check the status, and come back. That's friction that kills flow. /loop hands that job to Claude, so you can stay heads-down on code.

---

2. Voice Mode: Code by Talking to Your Terminal

Released: v2.1.74 (rolling out — currently ~5% of users as of March 26, 2026)

Voice mode is activated with the /voice command. It uses a push-to-talk model: hold the spacebar to speak, release to send. No always-on microphone, no accidental activations.

How to Enable It

/voice

Then hold spacebar and speak your instruction:

> "Refactor this function to use async/await and add proper error handling"

Release the spacebar — Claude transcribes and executes.

Customizing the Keybinding

You can change the push-to-talk key in ~/.claude/keybindings.json:

{
  "key": "meta+k",
  "command": "voice:pushToTalk"
}

Supported Languages (20 as of March 2026)

Previously available: English, Spanish, French, German, Italian, Portuguese, Japanese, Korean, Chinese, Hindi

New in March: Russian, Polish, Turkish, Dutch, Ukrainian, Greek, Czech, Danish, Swedish, Norwegian

Transcription is tuned for technical vocabulary — it correctly handles function names, library names, and common programming terms rather than mistranscribing them as random words.

Best Use Cases for Voice Mode

Fast iteration: Combine voice with the /effort low setting for rapid back-and-forth. Typing overhead is gone; you can fire instructions as fast as you can think them.

Long refactors: Describe a complex multi-step refactor verbally ("Extract this block into a separate utility function, rename the variables to camelCase, and add JSDoc comments") without hunting for the right typed phrasing.

Debugging by talking through the problem: Sometimes explaining a bug out loud helps you find it. With voice mode, that explanation directly becomes a Claude prompt.

---

3. Claude Opus 4.6: Now the Default Model + 1M Context Window

Opus 4.6 as Default (v2.1.68, March 4, 2026)

As of version 2.1.68, Claude Opus 4.6 replaced Opus 4.1 as the default model for all Max and Team plan users. Older models Opus 4 and Opus 4.1 were retired from the first-party API and users were automatically migrated.

What this means in practice: You get significantly stronger reasoning out of the box without changing anything. Opus 4.6 uses "medium" effort by default — balanced for everyday development tasks.

1M Token Context Window Goes GA (v2.1.75, March 13, 2026)

This is the headline upgrade. The 1 million token context window for Opus 4.6 is now generally available — no beta waitlist, no experimental flags.

Plan breakdown:

• Max / Team / Enterprise: 1M context is automatic, no extra charge, no settings to change
• Pro: Opt in by typing /extra-usage in Claude Code
• API (direct): Standard pricing at $5/$25 per million tokens (input/output). Long-context premium of $10/$37.50 per million tokens applies once input exceeds 200K tokens.

Why 1M Context Is a Game-Changer for Developers

The old 200K limit meant that large codebases triggered context compaction — Claude would summarize older parts of the session, potentially losing important details. With 1M tokens:

• Load entire repositories into context without worrying about compaction
• Maintain coherent multi-file refactors across hundreds of files
• Run long autonomous agent sessions that track their own history without drift
• Analyze entire codebases for security vulnerabilities in a single pass

For anyone running long-lived background agents or working on enterprise-scale codebases, this removes the single most frustrating constraint of the previous generation.

---

4. ultrathink: Maximum Reasoning on Demand

Re-introduced: v2.1.68 (March 2026)

The ultrathink keyword was originally part of earlier Claude releases, then removed. It's back in March 2026 as a way to activate high effort reasoning for a single turn — without permanently changing your effort setting.

How to Use ultrathink

Simply include the word ultrathink anywhere in your prompt:

ultrathink: design a distributed caching architecture for this API that handles 50k req/s

Or inline:

Review this authentication flow ultrathink and identify any security vulnerabilities

Claude treats this as a signal to allocate maximum reasoning budget to the response.

ultrathink vs /effort high

• /effort high — Sets high effort for the entire session going forward
• ultrathink — Activates high effort for one turn only, then reverts to your current effort level

Use ultrathink for specific hard problems (architecture decisions, complex debugging, security reviews) without burning extra compute on every routine query.

---

5. /effort Command: Fine-Grained Reasoning Control

Released: v2.1.76 (March 2026)

The /effort command replaces the old multi-level effort system with a simplified three-tier model:

The old "max" effort level has been removed — ultrathink in your prompt now serves that function for one-off maximum-effort turns.

Practical Workflow

# Set to low for a fast review session
/effort low

Back to high for a tricky bug
/effort high
ultrathink: why is this async function deadlocking in production

---

6. Computer Use: Claude Controls Your Mac Desktop

Status: Research Preview (v2.1.74, March 2026)

This is the most experimental feature in the batch, and potentially the most powerful. Computer Use lets Claude Code directly control your Mac — moving the mouse, clicking buttons, opening applications, browsing the web — without any manual setup or scripting.

What Claude Can Do in Computer Use Mode

• Open browsers, navigate to URLs, fill out forms
• Operate file managers and system apps
• Run terminal commands outside its session
• Interact with GUIs that have no API

When It Activates

Claude doesn't start randomly clicking things. Computer Use kicks in only when built-in tools are insufficient for a task — it's a fallback, not the default mode.

Example: You ask Claude to update a dependency in a GUI-only package manager. It'll attempt the task via terminal tools first; if that fails, it may switch to Computer Use to operate the application visually.

Note: This is still a Research Preview. Expect rough edges. Don't run it unattended on a production machine.

---

7. MCP Elicitation: Smarter Tool Calling

Released: v2.1.72 (March 2026)

MCP (Model Context Protocol) elicitation is a new mechanism that allows Claude to request additional information from MCP tools when it needs more context to proceed — instead of hallucinating parameters or failing silently.

In practice: if Claude is about to call an MCP tool but realizes it doesn't have a required piece of information (say, a database connection string or an API key), it will surface a prompt asking you to supply it. This makes complex multi-tool agent workflows significantly more reliable.

---

8. /color, /simplify, /batch: The Smaller QoL Wins

These three commands don't headline the update, but they add up.

/color — Visual Session Management (v2.1.75)

If you're running multiple Claude Code sessions in parallel (a common pattern with background agents), /color lets you assign a unique color to each session's prompt bar. When you have five terminal tabs open, instantly knowing which is which saves real time.

/color blue

/simplify and /batch (v2.1.63, February 28, 2026)

• /simplify — Refactors selected code to be more readable and concise
• /batch — Executes a sequence of operations in bulk on multiple targets

Both are particularly useful for cleaning up large codebases or applying consistent transformations across many files.

---

How to Update to the Latest Version

npm update -g @anthropic-ai/claude-code

Check your current version:

claude --version

As of March 26, 2026, the latest stable release is v2.1.76.

---

The Bigger Picture: Claude Code Is Becoming an Autonomous Platform

The March 2026 update batch signals a clear direction: Claude Code is evolving from a terminal-based coding assistant into a persistent, autonomous development platform.

• /loop gives it continuous monitoring capabilities
• Voice mode lowers the interaction barrier to zero
• 1M context lets it hold entire projects in memory
• Computer Use extends its reach beyond the terminal
• Background agents enable parallel workflows

The 74 releases in 52 days pace isn't accidental — Anthropic is in a shipping race with Cursor, GitHub Copilot, and Windsurf. Each of these features is a direct response to capability gaps developers have complained about publicly.

If you're already using Claude Code, update now and start with /loop and /effort — those will give you the most immediate workflow improvement. If you haven't tried it yet, the March 2026 release is the best entry point to date.

---

Frequently Asked Questions

Is voice mode available on all operating systems?

Currently Mac and Linux are confirmed. Windows support via WSL is being tested.

Does /loop use extra API credits?

Yes — each loop execution counts against your usage. Monitor your session usage with /status if you're on a limited plan.

Can I use 1M context on the Claude.ai web interface?

Not yet. The 1M context window is currently Claude Code-only (CLI). The web interface still uses the standard context window.

What happened to Opus 4 and Opus 4.1?

Both models were retired from Anthropic's first-party API in early March 2026. Existing users were automatically migrated to Opus 4.6. If you're on Bedrock, Vertex, or Microsoft Foundry, Opus 4.6 is now the default there too.

Does ultrathink work on all plans?

Yes, the keyword works on all plans — but actual effect scales with your plan's reasoning budget. Max plan users will see the most impact.