Claude Code March 2026 Upgrade: Ultrathink, Voice Mode, Skill Evals and More
Claude Code shipped four releases in a single week — v2.1.68 through v2.1.71. What looks like routine patch numbers is actually the biggest feature expansion since Claude Code launched in late 2025. Ultrathink returns, Voice Mode brings speech to the terminal, the Skill Creator gets a full eval framework, and /loop lets Claude Code schedule its own recurring tasks for the first time.
This article breaks down what changed in each release, why Anthropic is bundling these features right now, and what it means for dev teams and AI agencies using Claude Code in production.
Claude Code v2.1.68: Opus 4.6 Default, Ultrathink Returns
The March 4, 2026 release carries three changes that together reveal a clear strategy: Anthropic wants Claude Code to be faster and more efficient — without sacrificing quality when complex tasks demand it.
Opus 4.6 becomes the default model for Max and Team subscribers. The real change isn't the model — it's the effort level. The default shifts from "high" to "medium." Claude Code no longer burns maximum reasoning budget on every variable rename. It saves deep analysis for tasks that actually need it.
Ultrathink is back. The keyword was removed in an earlier release — a GitHub Issue (#19098) documented the quality regression that followed. Now typing "ultrathink" in a prompt sets the effort level to "high" for the next turn. Developers get control back: medium for daily work, ultrathink for the hard problems.
Opus 4 and 4.1 are removed from the first-party API. Users who had these models pinned are automatically migrated to Opus 4.6. Anthropic is cleaning house — a signal that the older model generation is no longer supported.
What This Means in Practice
For teams running Claude Code all day, the medium default cuts token costs significantly. At Context Studios, we noticed the difference immediately: routine tasks like refactoring, commit messages, and quick fixes run noticeably faster. For architectural decisions or complex debugging sessions, we type "ultrathink" — and get full reasoning depth back when we need it.
Claude Code Voice Mode: Talk Instead of Type
On March 3, 2026, Anthropic announced Voice Mode for Claude Code. The feature is currently available to about 5% of users and will roll out more broadly over the coming weeks — for Pro, Max, Team, and Enterprise plans.
Push-to-talk in the terminal: Hold spacebar, speak, release. The transcription appears at your cursor position. No always-listening mode, no wake word — deliberately designed for a terminal environment where accidental voice input is unwanted.
Hybrid input is the real feature. You can mix typed and spoken text in a single prompt. Type file paths and variable names precisely, speak the complex logic description freely. This solves a real problem: the gap between what you're thinking and what you type costs clarity and time.
Transcription tokens don't count against rate limits. Voice Mode costs nothing extra on any plan. Transcription processing runs separately from the model tokens used for Claude's responses.
Claude Code v2.1.69 extends speech-to-text support to 10 additional languages — Russian, Polish, Turkish, Dutch, Ukrainian, Greek, Czech, Danish, Swedish, and Norwegian — bringing the total to 20 languages. Claude Code v2.1.70 fixes Voice Mode issues on Windows and improves push-to-talk activation. Claude Code v2.1.71 makes the push-to-talk key configurable via voice:pushToTalk in keybindings.json and fixes a 5–8 second startup freeze when Voice Mode was active after system wake.
Skill Creator 2.0: Evals, Benchmarks, and A/B Testing for Agent Skills
On the same day as Voice Mode, Anthropic upgraded the Skill Creator — and for professional Claude Code users, this change might matter more than Voice Mode.
Agent Skills are markdown files that teach Claude Code to handle specific tasks in specific ways. The problem until now: you could write a skill file, but you couldn't systematically test whether it actually works — or whether a model update breaks it. As Anthropic explained in their announcement:
"Since launching Agent Skills last October, we've noticed that most authors are subject matter experts, not engineers. They know their workflows but don't have the tools to tell whether a skill still works with a new model, triggers when it should, or if it actually improved after an edit."
— Anthropic, "Improving Skill-Creator: Test, Measure, and Refine Agent Skills" (March 2026)
The updated Skill Creator solves this with three features:
1. Eval generation without code. Define test prompts, describe the expected outcome, and the Skill Creator checks whether the result matches expectations. This is software testing for skill authors who aren't engineers — and according to Anthropic, that's the majority of skill authors.
2. Benchmark mode. A standardized run across all evals that tracks pass rate, token usage, and elapsed time. This enables comparisons across model updates: does my skill still perform the same after a Claude update?
3. Multi-agent evals with comparator agents. Evals run in parallel in isolated contexts — no context bleed between tests. Comparator agents enable A/B comparisons: two skill versions, or skill vs. no skill, evaluated without knowing which output is which.
Anthropic distinguishes two categories of skills: Capability uplift skills that teach Claude something the base model can't (consistently) do, and Encoded preference skills that orchestrate existing capabilities in a team-specific workflow. Evals deliver value for both — but for different reasons. Capability skills can become obsolete as models improve. Evals show when that happens. Preference skills are more durable but only as valuable as their fidelity to the actual workflow.
Why This Matters for AI Studios
At Context Studios, we run over a dozen Agent Skills — from content pipelines to SEO audits to code review workflows. Until now, every model update was a risk: does the skill still work? Does it even trigger? With evals and benchmark mode, we can verify before deployment instead of waiting for production failures. That's the difference between "it seems to work" and "we know it works."
Claude Code v2.1.69: /claude-api Skill, Extended Hooks, and 60+ Bug Fixes
Claude Code v2.1.69 is by far the most extensive patch release in recent months. Beyond the Voice STT expansion to 20 languages, it brings:
- The
/claude-apiskill — a built-in guide for building applications with the Claude API and Anthropic SDK. This lowers the barrier for developers moving from CLI to API. ${CLAUDE_SKILL_DIR}variable — skills can now reference their own directory. This makes skills more portable and eliminates hardcoded paths.- InstructionsLoaded hook event — fires when CLAUDE.md or
.claude/rules/*.mdfiles are loaded. Enables automation around the skill lifecycle. /reload-pluginscommand — activate plugin changes without restarting Claude Code.- Custom Remote Control names —
/remote-control My Projectsets a custom session title visible in claude.ai/code. - Security fix: Nested skill discovery could load skills from gitignored directories like
node_modules— now patched.
The release includes over 60 bug fixes, including memory leaks in long sessions, Windows stability improvements, and a critical fix to the MCP server trust dialog.
Claude Code v2.1.70: VS Code Gets an Upgrade
The March 6 release is a stability and polish release with a strong VS Code focus:
- Spark icon in the VS Code Activity Bar — lists all Claude Code sessions, which open as full editors.
- Markdown document view for plans — with comment support for feedback directly in VS Code.
- Native MCP server management —
/mcpin the chat panel enables, disables, and reconnects MCP servers without switching to the terminal. - 300× reduced Remote Control polling rate — from 1–2 seconds to once per 10 minutes while connected. Reconnection remains instant.
- Prompt input re-renders reduced by 74% — noticeably smoother UI.
Claude Code v2.1.71: /loop and Cron Scheduling
The latest release from March 7 moves Claude Code one step closer to an autonomous agent:
/loopcommand — recurring prompts or slash commands at fixed intervals:/loop 5m check the deploy. This enables continuous monitoring directly from a Claude Code session.- Cron scheduling tools — for the first time, Claude Code can schedule its own recurring tasks within a session.
- Configurable push-to-talk key —
voice:pushToTalkin keybindings.json, e.g.,meta+kfor zero typing interference. - Bridge session reconnection — reconnects after laptop wake in seconds instead of up to 10 minutes.
The Big Picture: Claude Code Is Becoming a Team Platform
Four releases in one week isn't coincidence. Anthropic is pursuing a recognizable strategy: Claude Code is evolving from an individual developer tool into a team-capable, persistent coding platform.
The evidence:
- Voice Mode lowers the barrier to entry — developers who think faster than they type can now work more productively.
- Skill Creator with evals brings software engineering practices to skill development — essential for teams maintaining skills across model updates.
- Remote Control with custom names and VS Code MCP management make multi-session work more manageable.
/loopand cron scheduling transform Claude Code from a reactive tool into a proactive agent that monitors and acts autonomously.- The revenue numbers back the direction: Claude Code's run-rate revenue exceeded $2.5 billion according to Anthropic's Series G announcement — more than doubling since the beginning of 2026.
As Kate Jensen, Anthropic's Head of Americas, stated at the company's Enterprise Agents briefing:
"In 2025 Claude transformed how developers work, and in 2026 it will do the same for knowledge work. The magic behind Claude Code is simple. When you can delegate hard challenges, you can focus on the work that actually matters."
— Kate Jensen, Head of Americas at Anthropic, "Briefing: Enterprise Agents" (March 2026)
For AI development studios like Context Studios, this means Claude Code is no longer just a tool that makes developers more productive. It's becoming the platform where agent-based workflows run — with testable skills, voice-controlled interaction, and built-in automation.
FAQ
What is Claude Code Ultrathink and how do you activate it?
Claude Code Ultrathink is a keyword that sets the reasoning effort level to "high" for the next turn. Simply type "ultrathink" in your prompt, and Claude Code will analyze the task with maximum reasoning depth. It was reintroduced in v2.1.68 on March 4, 2026, after users documented quality regressions following its earlier removal. The default effort level is now "medium" — Ultrathink gives you full analysis on demand.
How does Voice Mode work in Claude Code?
Voice Mode uses push-to-talk: hold spacebar, speak, release — the transcription appears at your cursor position. You can mix typed and spoken text in a single prompt. Since March 3, 2026, it's rolling out gradually starting with 5% of users. Available on Pro, Max, Team, and Enterprise plans. Toggle it on with /voice in the terminal.
What changes with the Skill Creator 2.0?
The Skill Creator now supports eval generation, benchmarking, and A/B testing via comparator agents. Skill authors can define tests without writing code and track results across model updates. Multi-agent evals run in parallel in isolated contexts. Available as a plugin for Claude Code and natively in Claude.ai/Cowork.
What is the /loop command in Claude Code v2.1.71?
The /loop command runs prompts or slash commands at recurring intervals — for example, /loop 5m check the deploy. Combined with the new cron scheduling tools, it enables continuous monitoring and automation directly from a Claude Code session, without external tools.
Which languages does Claude Code Voice Mode support?
Claude Code Voice Mode supports 20 languages for speech-to-text transcription as of v2.1.69. Beyond English, it added Russian, Polish, Turkish, Dutch, Ukrainian, Greek, Czech, Danish, Swedish, and Norwegian. The push-to-talk key is freely configurable via keybindings.json since v2.1.71.