Claude Products in 2026: What Changed Since Launch
The three Claude products got serious upgrades in 2026 — new models, bigger context, cloud agents, and a design studio. Here's the current state.
I wrote about Claude’s three products back in Episode 1. That was the initial breakdown: Claude.ai for conversations, Cowork for desktop automation, Code for engineering. The bones haven’t changed. But the muscle on top? Completely different animal.
Here’s what’s new and why it matters.
The Model Lineup: One Family, Three Sizes
All three products run the same model family. The specific models as of May 2026:
- Opus 4.7 — deepest reasoning, handles the hardest problems
- Sonnet 4.6 — the workhorse, now with a 100M token context window
- Haiku 4.5 — fast and cheap, 90% of Sonnet’s capability at a third of the cost
That 100M token context on Sonnet is the headline. For perspective: a full-length novel is about 100K tokens. Sonnet can now hold roughly 1,000 novels worth of context in a single conversation. In practice, this means you can feed it an entire codebase, an entire documentation set, or months of conversation history without hitting the wall.
The model you use depends on the product. Claude.ai defaults to Sonnet but lets you switch. Cowork uses Sonnet for most tasks, Opus for hard ones. Claude Code gives you the most control — you pick the model per session.
Claude.ai: Projects + Design Studio
The web interface got two upgrades that changed how I use it.
Projects turned Claude.ai from a stateless chatbot into something with institutional memory. A project is a persistent workspace with custom instructions and a RAG knowledge base. Upload your company docs, product specs, market research — Claude references them automatically in every conversation inside that project.
I have a project called “Etsy Research” with 40+ competitor reports and category data loaded into it. When I ask a question about pricing trends, Claude pulls from those documents without me pasting anything. Sounds simple. Saves 15 minutes per session when you’re doing it 10 times a day.
Claude Design launched April 17, 2026. This one surprised me. You describe what you want in plain English — “a 5-slide pitch deck for an AI SaaS product, dark theme, modern typography” — and it generates the actual visual content. Not a wireframe. Not a mockup. A finished presentation with real layout, colors, and typography.
Export options: PDF, PPTX, or hand it to Claude Code to convert into working frontend code. You can also push designs directly to Canva for further editing.
Best for: research, writing, brainstorming, visual content creation. Zero technical background required.
Claude Cowork: Your Desktop Is the API
Cowork went GA on April 9, 2026 for macOS and Windows simultaneously. The positioning is clear: it’s not a chatbot that lives in a window. It’s an agent that operates your entire desktop.
Computer Use is the core feature. Cowork physically controls your mouse and keyboard. It clicks buttons, fills forms, switches between apps, drags files. “Take these 12 Excel spreadsheets, merge them into a summary report, and email it to my team” — it does that. You watch it happen in real time, or you walk away and come back to a finished result.
355 app connectors across 33 categories. That number was 50 six months ago. The coverage now includes Slack, Notion, Google Workspace (Docs, Sheets, Slides, Gmail), GitHub, Linear, Jira, Figma, and dozens more. Plus MCP protocol support, which means any custom integration you build plugs in natively.
MCP integration deserves its own line. Model Context Protocol is Anthropic’s open standard for connecting AI to external tools. Cowork supports it natively. If you’ve built an MCP server for your internal database, your CRM, your inventory system — Cowork can talk to it without any special adapter.
Best for: workflow automation, cross-app tasks, anything that involves clicking through multiple tools in sequence.
Claude Code: The Engineering Platform
This is the one I use 8+ hours a day. The upgrades here are structural, not cosmetic.
SWE-bench 87.6%. Per Vellum’s benchmark data, Claude Code holds the top score on the industry-standard coding benchmark. Gemini 3.1 Pro sits at 80.6%. Seven points might not sound like much. In practice, it’s the difference between an agent that needs hand-holding on complex tasks and one that ships features end-to-end.
Routines launched April 14, 2026. Before this, Claude Code only worked while your terminal was open. Routines run on Anthropic’s cloud infrastructure — close your laptop, doesn’t matter. Schedule them hourly, daily, weekly. Trigger them via HTTP API or GitHub webhooks.
Use cases I actually run: a daily routine that checks my GitHub repos for dependency updates and opens PRs. A weekly routine that audits my Cloudflare Pages sites for broken links. A routine triggered by new GitHub issues that drafts a response and labels them.
Pro plan: 5 routine runs per day. Max plan: 15 per day. The limit is per-run, not per-routine — one routine executing once counts as one run.
Skills ecosystem. Skills are community-built workflow packages that Claude Code auto-invokes when relevant. I’ve built skills for image generation (controls Chrome via CDP to call Gemini), GitHub operations (one sentence manages repos), and browser automation. The skill format is simple: a markdown file with trigger conditions, a workflow, and constraints.
The difference between 2025 Claude Code and 2026 Claude Code: it went from a coding assistant to a software engineering platform. Write code, test it, deploy it, monitor it, fix it when it breaks — all from the terminal.
Best for: software engineering, automation pipelines, AI agent orchestration. Requires terminal comfort.
The Decision Framework
Forget feature comparisons. Ask yourself one question: what’s the output?
Need visual output? Claude.ai. Presentations, documents, research reports, design assets. Claude Design handles the visual side. Projects handle the knowledge side.
Need app automation? Claude Cowork. If the task involves clicking through 3+ apps in sequence, Cowork wins. Computer Use plus 355 connectors means you describe the workflow once and walk away.
Need code? Claude Code. If the output is a git commit, a deployed service, an automation script, or an AI agent — there’s no substitute.
Need all three? That’s the point. One Max subscription ($200/month) covers all three products. Design a prototype in Claude.ai, automate the deployment workflow in Cowork, build the actual product in Claude Code. They share models, they share context (via MCP), and increasingly they share projects.
The natural progression hasn’t changed since Episode 1. But the ceiling on each product is dramatically higher than it was six months ago.