Anthropic Ships Agent OS: Opus 4.7, Claude Design & Cowork

Claude is now an AI agent OS — and Anthropic shipped the features to prove it. In four months, Anthropic quietly turned Claude from a text chatbot into something closer to a full operating system for knowledge work. Since January 2026, the company shipped three flagship models — Opus 4.7, Sonnet 4.6, and Haiku 4.5 — plus visual design tools, a local agent runtime, computer control, and enterprise-grade compliance features. This is not iteration. It is a platform pivot.

From Chat Window to Agent Control Tower

The clearest way to understand the scale of change: six months ago, Claude answered questions. Today it can open your files, click through browser tabs, design a slide deck, schedule recurring tasks, and remember every conversation you have ever had — without you touching a keyboard for the task itself.

The convergence of three specific features defines this sprint:

• Claude Design (April 17, 2026) — create slides, prototypes, and one-pagers visually inside Claude, without switching to Figma or Canva
• Computer Use (research preview) — Claude clicks, navigates, and operates tools on your behalf, turning the model into an operator rather than a text advisor
• Cowork (general availability April 9, 2026) — a local agent runtime (a program that executes tasks on your own machine inside an isolated virtual environment, so it cannot damage your main system) available on both macOS and Windows

Previously, each of these required separate tools: Figma for design, Zapier for automation, GitHub Copilot for code. Anthropic is betting users will consolidate onto one platform if it covers all three well enough.

Three Model Tiers, Three User Profiles

Anthropic shipped three distinct model tiers in 2026, each aimed at a different use case and cost point. Understanding which model does what matters when choosing a plan — and when planning enterprise deployments.

Opus 4.7 — Flagship With Upgraded Vision

Launched April 16, 2026, Opus 4.7 improves software engineering performance and adds support for higher-resolution image understanding. Claude can now review a screenshot of broken code, a wireframe, or a filled spreadsheet — and give precise feedback based on what it actually sees, not a text description you typed about it. Think of the difference between describing a painting over the phone versus handing someone the canvas directly.

Sonnet 4.6 — The 1-Million Token Context Window

Launched February 17, 2026, Sonnet 4.6 introduced a 1-million token context window (currently in beta — stable for early adopters but not yet guaranteed for all customers). A token (the unit AI models use to measure text length, roughly three-quarters of a word) means 1M tokens equals approximately 750,000 words: six full-length novels in a single conversation, without the model losing track of details from page one.

Anthropic claims Sonnet 4.6 matches or exceeds Opus-level capabilities on coding and agent tasks at lower cost. If validated at scale, that shifts enterprise pricing math: flagship-level output without the flagship price tag.

Haiku 4.5 — Fast, Cheap, and Now Flagship-Accurate

Haiku 4.5 (launched October 2025) now matches Sonnet 4's performance on coding, computer use, and agent tasks — at faster speed and lower per-token cost (the unit cost AI providers charge per chunk of text processed). This is the high-volume tier: useful when your application needs thousands of model calls per minute and does not require Opus-level reasoning on every single one.

Claude Design: AI Gets a Visual Layer

Claude Design, part of Anthropic Labs (Anthropic's experimental product division, separate from the main Claude product line), launched April 17, 2026. It lets users collaborate with Claude to produce visual outputs — presentations, prototypes, single-page documents — without leaving the chat interface.

This is Anthropic's first native visual creation product. OpenAI has DALL-E for image generation (creating pictures from text descriptions) but no Canvas-style design environment. Figma and Canva remain standalone tools. Claude Design's core pitch is integration: your research, your draft, and your slide deck all live in one conversation thread. As of March 12, 2026, interactive charts, diagrams, and visualizations also render inline directly in Claude's chat responses — no export step required.

Computer Use and Cowork: Claude Takes the Wheel

Computer Use is now available as a research preview (a limited early-access release without full stability guarantees) for Pro and Max subscribers. In practice: Claude can open files on your system, run developer tools, click through browser interfaces, and complete tasks you would otherwise handle manually.

The practical shift is significant. Instead of copying Claude's output into a spreadsheet, Claude opens the spreadsheet directly. Instead of describing a bug, it runs the code, reads the error message, and fixes it in context. The model stops being an advisor and becomes an operator.

Cowork serves a related but distinct purpose. It is a local agent runtime — running on your own computer inside an isolated virtual machine (a self-contained software environment separated from your main system so agent actions cannot cause unintended damage) — with access to your files and third-party tool connections. Capabilities as of April 2026:

• Scheduled recurring and on-demand tasks, with admin controls for Team and Enterprise organizations
• Integration with Slack, Gmail, Google Calendar, Google Docs, and GitHub via Claude in Chrome (beta)
• macOS support since January 2026; Windows support added April 2026
• Local execution keeps data on-device — critical for organizations with data residency or compliance requirements

That last point is concrete. Cloud-dependent agent platforms send task data to external servers. Cowork running locally in an isolated VM gives organizations a genuine on-premise option — important for legal, healthcare, and government teams where data leaving the building is a compliance issue. Read more about setting up AI automation workflows for your team.

Enterprise Goes Self-Serve — HIPAA, Analytics, and Role Controls

The enterprise feature push is the least-discussed but commercially most significant part of this sprint. Three structural changes stand out:

• Self-serve Enterprise plans (February 12, 2026) — organizations can sign up without a sales conversation, with a single seat type covering Claude, Claude Code, and Cowork access
• Role-based access controls — admins organize users into groups and restrict which Claude capabilities each department can access (marketing gets different tool access than engineering, for example)
• Analytics API (February 13, 2026) — programmatic access (meaning IT teams can query usage data via code rather than clicking through a dashboard) to aggregated usage and engagement data per organization and per day
• HIPAA-ready Enterprise plans — health and fitness data analysis is live for US Pro/Max users; Enterprise plans offer full HIPAA compliance (the US federal standard for protecting medical records) for healthcare organizations handling sensitive patient data
• Agent Skills marketplace (December 18, 2025) — an open standard for organization-wide skill management and a growing directory of partner-built integrations

Self-serve Enterprise removes a meaningful adoption barrier. The traditional AI enterprise sales motion — multi-month pilots, custom contracts, procurement cycles — is too slow for teams that want to deploy this quarter. Removing the sales requirement means faster time-to-deployment and a fundamentally different go-to-market motion for Anthropic.

The Limits That Matter Before You Deploy

Before treating this as a finished platform, the preview status of several features has real planning implications for IT teams and early adopters:

• The 1M token context window in Sonnet 4.6 is still in beta — production stability is not yet guaranteed for all customers
• Computer Use is a research preview — Pro and Max plans only, not generally available across all tiers
• Claude in Chrome is in beta — not fully stable across all workflow types
• Health data features are US-only on Pro/Max plans — international users and free-tier accounts are excluded entirely
• Windows Cowork support arrived three months after macOS (January to April) — expect staggered rollouts for future capabilities
• Persistent Cowork thread control remains in research preview as of March 2026, with no confirmed general availability date

None of these are dealbreakers — they reflect the pattern of a company shipping fast to early adopters and then expanding access. But if you are an enterprise IT admin building a deployment plan around any beta feature, budget for general availability timelines that may shift by weeks.

Anthropic's Bet: One Model Stack to Replace Three Tools

The post-Easter 2026 timing is deliberate. Google has Gemini-powered Workspace integrations. OpenAI has Canvas, Operator, and a $200/month Pro tier. Microsoft Copilot is embedded in every Office product. Anthropic's answer: not one flagship model, but a layered model stack — Opus 4.7 for complex reasoning, Sonnet 4.6 for high-context tasks, Haiku 4.5 for fast bulk processing — plus agents, visual design tools, and enterprise controls, all converging on a single platform.

Three model tiers give Anthropic price-to-performance coverage that a single flagship cannot provide. An organization can deploy Haiku 4.5 for routine document classification and Opus 4.7 for complex legal analysis — inside the same platform, under the same compliance umbrella, with the same role-based access controls and usage audit trail. That is a compelling enterprise pitch that none of the alternatives currently match end-to-end.

If you are already on a Pro or Max plan, the most actionable experiment is this: set up one recurring Cowork task for something you currently do manually every week. The combination of scheduled AI automation and local file access is the feature pair most likely to replace a paid tool you are already using separately. Try it for two weeks — that is where this platform sprint is pointing, and the clearest way to see whether the convergence is real for your specific workflow. Our AI automation quick-start guide walks through the first Cowork configuration step by step.

Related Content — Get Started | Guides | More News