Claude Code has been quietly evolving faster than most people realize. While the AI discourse stays fixated on model releases and benchmark wars, Anthropic has been shipping daily updates to Claude Code — adding new primitives, expanding access, and building toward something that looks less like a coding assistant and more like a programmable agent layer for serious work. If you set it up in 2025 and haven’t revisited it, you’re missing a lot. Here’s what’s actually changed, what’s worth your attention, and how to use it well in 2026.
What Claude Code Actually Is in 2026
Claude Code started as a command-line tool for developers who wanted an AI that could read their codebase, write code, run terminal commands, and iterate without constant hand-holding. That’s still the core. But the scope has expanded considerably.
It’s now included in every Team plan standard seat — no add-on required. That alone changed who’s using it. It’s also the engine behind Claude Cowork, Anthropic’s desktop app for knowledge workers, which was built using Claude Code in ten days by Anthropic’s own team. That’s not a marketing claim — it’s a signal about how seriously Anthropic is using its own tooling internally. Anthropic engineers reportedly use Claude for around 60% of their work and are shipping 60 to 100 internal releases per day. Whether you read that as inspiring or alarming probably depends on your role.
The underlying models have also changed. Claude Opus 4.6 is now the default for Max, Team, and Enterprise users in Claude Code, with a 1 million token context window. That’s not a novelty — it means Claude Code can hold your entire large codebase, test suite, documentation, and conversation history in a single context window without truncating or losing track of earlier decisions. Anthropic’s own Frontier Red Team used Opus 4.6 to find more than 500 vulnerabilities in production open-source code. That’s a concrete demonstration of what the model can do with deep, sustained focus on a complex codebase.
Claude Sonnet 4.6 launched February 17, 2026, at the same price as 4.5 with better performance and improved agentic search. It also supports a 1 million token context window in beta. Opus 4 and 4.1 have been deprecated from the model selector, so if you have old scripts or configurations pointing to those, update them.
The Skills API: Finally, Reusable Agent Behaviors
The most underrated feature Claude Code shipped recently is the Skills API, also called Agent Skills. The idea is simple but it solves a real problem: if you’ve spent time getting Claude Code to behave exactly the way you want for a specific task — writing commit messages, generating slide decks, parsing a particular data format — you used to have to re-explain that every time or rely on a bloated system prompt. Skills changes that.
You organize skills into folders with a SKILL.md file that describes what the skill does, when to use it, and how to invoke it. Claude Code reads these and applies the right skill automatically when relevant, or you can call them explicitly.
Anthropic ships pre-built skills for PPTX, XLSX, DOCX, and PDF — which matters because a huge amount of real business work involves those formats. You can now have Claude Code generate a structured PowerPoint presentation, edit Excel pivot tables and conditional formatting, or produce a formatted Word document as part of an automated workflow, without writing custom parser code for each format.
For developers building on top of Claude Code, this is the primitive that enables you to package repeatable agent behaviors and share them across a team. Think of it as functions, but for agent behavior. A legal team could have a contract-review skill. A finance team could have an earnings-analysis skill. A dev team could have a PR-review skill that checks against their specific conventions. The organizational implications here are significant, even if the interface is still relatively low-level.
Channels and Bare: The Automation Flags Worth Knowing
Two newer flags in Claude Code are aimed at different ends of the automation spectrum, and they’re easy to miss if you’re not watching the release notes.
–channels is a permission relay system currently in research preview. It lets Claude Code communicate across different contexts or agent threads — essentially enabling multi-agent coordination where one Claude Code instance can hand off or relay information to another. The practical use case is building pipelines where different agents handle different parts of a workflow with appropriate permission boundaries between them. This is still early and in research preview, so treat it as something to experiment with rather than build production systems on today. But the direction is clear: Anthropic is building toward orchestration, not just single-agent operation.
–bare is more immediately useful for anyone doing scripted automation. It strips Claude Code’s output down to raw responses without the surrounding interface elements, making it much easier to pipe Claude Code into shell scripts, CI pipelines, or other automated systems. If you’ve tried to use Claude Code programmatically and found the output format annoying to parse, –bare is what you wanted.
Combined, these two flags sketch out the near future of Claude Code: a composable automation layer that can be wired into existing systems, not just a terminal tool you interact with manually.
Claude Cowork and the “Vibe Working” Frame
Claude Cowork launched in research preview at the end of January 2026. It’s a macOS desktop app that runs Claude in an isolated VM on your local machine, with full access to local files and MCP integrations. Scott White, Anthropic’s Head of Product for Enterprise, described the intent as “transitioning into vibe working” — knowledge workers directing AI instead of writing code themselves.
The parallel to “vibe coding” is intentional. Andrej Karpathy popularized the term vibe coding to describe a mode of software development where you describe what you want in natural language and the model handles implementation. Cowork extends that frame to non-developers — the analyst who needs a financial model built, the operations lead who wants to automate a workflow, the HR director building a document pipeline. You don’t need to understand Python to direct an agent that writes Python.
Cowork includes domain-specific plugins for legal, financial analysis, HR, engineering, and operations. It supports persistent agent threads for Pro and Max users across mobile and desktop. The “persistent thread” piece is worth noting — it means the agent maintains context across sessions, so you’re not starting from scratch every time you open it. That’s the difference between a tool you use occasionally and a collaborator that builds up context about how you work.
Cowork is still in research preview as of March 2026, which means limited access and rough edges. But it’s a direct signal of where Anthropic thinks the product is going: away from developer-only and toward a general-purpose AI work environment.
Claude in Your Existing Tools: Browser, Excel, PowerPoint
If you’re not in a terminal or a dedicated AI app, Claude is also showing up in the tools people already use every day.
Claude in Chrome is a browser extension that can read console errors, DOM structure, and network requests. For developers debugging a web app, this is genuinely useful — instead of copying error messages into a chat window, Claude can see the actual browser conte