Between February 3 and March 24, 2026, Anthropic shipped 74 documented product releases across four product surfaces simultaneously. That works out to roughly 1.4 releases per day, sustained for almost two months, without any single team waiting for another to finish.
The breakdown: 28 releases in Claude Code, 15 in Cowork, 18 in API and infrastructure, and 13 across models and core platform. No comparable AI lab has publicly demonstrated this kind of coordinated shipping velocity across this many surfaces at once.
What follows is not a comprehensive list — that would require a separate article. These are the 21 features that changed how Claude actually works, with enough context to understand why they matter.
Models & Core Platform
1. Claude Opus 4.6 (February 5)
The biggest single release of the period. Opus 4.6 launched with a 1-million-token context window in beta, 128,000 max output tokens per response (double the previous cap), and topped the METR time-horizon benchmark at 14.5 hours — a tenfold increase over the previous year. For the first time, a single Claude response can contain an entire codebase module or a 50,000-word research document without truncation. On the Finance Agent benchmark, it took the top spot across all models.
2. Claude Sonnet 4.6 (February 17)
A full capability upgrade across coding, computer use, long-context reasoning, agent planning, and design. Sonnet 4.6 matches near-Opus performance on SWE-bench (79.6%) at roughly a fifth of the cost. For most production agentic workflows, this is now the rational default — the model most developers actually run. It also introduced the 1M token context window in beta.
3. Fast Mode for Opus
A speed-optimized response mode for Claude Code sessions using Opus. When the task is time-sensitive and the code is straightforward, Fast Mode trades depth for latency — bringing Opus-class reasoning at Sonnet-class response times for simpler operations. Not the default; opt in for sessions where speed matters more than maximum depth.
4. Inline Visualizations in Chat
Claude can now generate charts, diagrams, and other visual content directly in chat responses, rendered inline alongside text. A data question can now return an answer and a chart in the same turn — without separate tools or artifacts.
Claude Code
5. Agent Teams (Research Preview, February 5)
A lead agent can spawn multiple sub-agents that work in parallel on different parts of a codebase and coordinate with each other. One handles tests, one handles implementation, one reviews. For large, cross-file tasks this changes what Claude Code can accomplish in a single session. Currently in research preview.
6. Session Sharing (February 3)
Share a complete Claude Code session — the full conversation, tool calls, file edits — with a teammate via link. The recipient sees exactly what happened: every decision, every command, every output. A concrete paper trail for async code review and collaborative debugging across time zones.
7. /insights Command (February 4)
Reads your session history from the past month and generates a structured analysis: which projects you worked on, how you used Claude Code, and concrete suggestions for improving your workflow. It shows you where your own habits are creating friction — a rare kind of productivity feature.
8. /rewind with Summarization (February 6)
When you rewind a conversation (ESC twice or /rewind), Claude now summarizes the discarded section instead of silently dropping it. You can take a different path and carry the learnings back in time without losing the context of the dead end. In long debugging sessions, this matters.
9. /debug Skill (February 5)
A dedicated debugging workflow triggered by /debug. Claude reads your session's debug logs and helps you trace failures — why a hook didn't trigger, why a tool call returned unexpected results, why a build step broke. Not a general troubleshooter, but a focused diagnostic skill that understands the structure of a Claude Code session.
10. Desktop App: Preview + CI/CD (February 20)
The Claude Code desktop app can now preview your running application directly in the interface and handle CI/CD failures and pull request workflows in the background. Ask Claude to fix a failing CI job while you keep working — it monitors the status, makes targeted changes, and reports back when the pipeline passes.
11. Git Worktrees (February 20)
Built-in worktree support lets multiple Claude Code agents work on parallel branches without switching context or stepping on each other. For multi-branch workflows — maintaining a fix branch while developing a feature branch simultaneously — this removes the consistent friction of manual git management between sessions.
12. Claude Code Security (February 20, Limited Preview)
A reasoning-based vulnerability scanner that traces data flows across an entire codebase rather than matching against a library of known-bad patterns. Anthropic's own testing against well-reviewed open-source projects found over 500 previously undetected vulnerabilities, including bugs that had survived years of review. False positive rate under 5%, compared to 30–60% for traditional static analysis. Currently in limited research preview.
13. Remote Control (February 25)
A live session mirror that lets you monitor and control a running Claude Code session from your smartphone or any browser. The session runs locally; the phone is the remote control. Sleep-resilient, outbound-HTTPS-only, no port forwarding required. Covered in detail in a previous post on this blog.
14. Figma MCP Integration (February 18)
Push Claude Code outputs directly into Figma. Build a working prototype in code, then send it to a Figma canvas to explore multiple design versions. The reverse is also possible with the Stitch MCP (covered separately on this blog): design in Stitch, pull the result into Claude Code as working code.
15. Claude Code Channels (March 20)
Control a Claude Code session asynchronously via Telegram or Discord. You send a message, Claude works on the codebase locally, and reports back when done or when it needs input. Code stays on your machine; the messaging apps are the remote control layer. Covered in detail in a previous post on this blog.
16. PowerShell Support on Windows (Late March, Preview)
An opt-in preview adding a PowerShell tool to Claude Code on Windows, alongside the existing Bash tool. A prerequisite for Computer Use on Windows and removes the friction for Windows developers without a WSL environment.
Claude Cowork
17. Cowork Launched (January 12 — just outside the window)
Technically before February 1, but impossible to omit: Cowork is a file-and-folder-based desktop AI assistant for knowledge workers who don't live in terminals. You grant Claude access to a directory, describe a task, and Claude reads existing files, executes the workflow, and produces deliverables. Launched with Slack, Gmail, Google Calendar, and Google Drive connectors.
18. Dispatch: Persistent Mobile Agent Thread (Mid-March)
A persistent agent thread accessible from phone or desktop, letting you assign tasks to Claude while away from your computer. Paired with Computer Use: assign the task from your iPhone, return to finished work on the desktop. One of the more concrete materializations of the "ambient AI" concept.
19. Scheduled Tasks in Cowork (Late March)
Create recurring and on-demand tasks that run automatically — daily morning summaries, weekly report generation, recurring data pulls, PR monitoring. Claude Code can now function as a lightweight cron service for workflows involving file reads and document generation.
20. Plugin Marketplace for Team/Enterprise (Late March)
A curated marketplace of Cowork plugins with admin controls for Team and Enterprise plans. Admins can approve, restrict, or allowlist specific plugins for their organization — directly addressing the enterprise concern about shadow IT.
API & Infrastructure
21. Enterprise Analytics API (March)
Programmatic access to usage and engagement data for Claude and Claude Code Remote usage within an organization — aggregated per organization, per day. The observability layer that was missing: actual usage numbers, cost per team, code acceptance rates. Enables real cost attribution and anomaly alerting — prerequisites for responsible deployment at scale.
What the Velocity Itself Says
There is a structural reason the pace has accelerated: Anthropic's engineers reportedly use Claude for roughly 60% of their own work, up from 28% a year ago. The team ships 60–100 internal releases per day. When the tools you're building make you faster at building tools, the release cadence compresses. This is not a metaphor — it is a feedback loop with a measurable effect on external shipping velocity.
The pattern across these 74 releases is also readable. Almost every feature either extends the reach of Claude's autonomous capabilities (Agent Teams, Channels, Dispatch, Worktrees, Computer Use) or deepens the quality of what happens at each step (Opus 4.6, Code Security, /rewind with summarization, /debug). There are very few cosmetic features. The roadmap appears to be: expand the surface area of autonomous action, then deepen capability within each surface.
For developers building on Claude, this creates a compounding opportunity — and a compounding risk of falling behind. The practical response is not to track every release, but to build a deliberate review habit: quarterly evaluation of which new capabilities apply to your workflows, and which to skip.