Three announcements in four days.
On Tuesday 14 April 2026, Anthropic rebuilt the Claude Code desktop app — a new sidebar, parallel Sessions, scheduled Routines, an integrated terminal, and a rebuilt diff viewer, all in one post (Anthropic blog, 14 Apr). On Thursday, Opus 4.7 went generally available (Anthropic, 16 Apr). On Friday, Claude Design shipped — a whole new surface at claude.ai/design where you describe what you want and Claude draws it for you on a live canvas (Anthropic Labs, 17 Apr).
None of those were small. Put together, the shape of Claude feels different. The engine got more capable, the builder got roomier, the automation layer grew, and a new creative surface appeared that doesn’t look like any of the others.
This piece is the tour. What each thing is, how it actually feels to use, what it costs you, and the honest caveats that aren’t in the marketing.
The headliner: Claude Design
Claude Design is the one that made the most noise, because it’s the most visible. It lives at claude.ai/design and it’s available to anyone on a Pro, Max, Team, or Enterprise plan — no separate subscription, no beta waitlist.
What it is, in one sentence: a room where you talk to Claude and watch it draw.
Chat on the left, live canvas on the right. You describe something — a landing page for your course, a three-slide pitch deck, a prototype of an app you’re thinking about, a one-pager for a client — and Claude produces it on the canvas, rendered, interactive, working. Not a mockup. Not a static image. A real thing you can click through.
The outputs Anthropic officially lists are:
- Interactive prototypes. Screens that scroll, forms that type, state that changes. The kind of thing you’d usually have to hand to a developer to see.
- Product wireframes and mockups. Feature flows sketched out and passable to Claude Code for actual implementation later.
- Pitch decks and presentations. Rough outline in, polished on-brand deck out.
- Landing pages and marketing collateral. Hero sections, pricing tables, social assets.
- Frontier experiments. Code-powered prototypes with voice, video, shaders, 3D, embedded AI.
No image generation. This is the first thing to know. Claude Design will not make you a photorealistic hero image of a woman in a studio. It will draw an SVG icon, an illustration, a diagram, a chart — anything vector — but not photography and not product renders. For those, you still leave and go to Flora or Midjourney.
The trick that makes it feel different
Most design tools ask you to start with a blank page. Claude Design does something quieter. During setup, it reads your codebase or your existing design files and builds a design system for your team — colours, typography, components, spacing. Then every time you generate something, it uses that system.
So when you say “make me a landing page for my course,” you don’t also have to say “in navy and cream, sans-serif, rounded corners, eight-column grid.” It already knows. It reads it off your repo.
For anyone who has ever tried to get an AI to produce something that matches a brand, this is the part that matters. Not the prompting. The reading.
How you actually use it
Start a project. Upload what you have — screenshots, a DOCX brief, a PPTX pitch, a spreadsheet of copy, a link to your existing site. Point it at a codebase if you have one.
Then describe what you want in plain language. The same voice you’d use talking to a friend:
Make me a landing page for my course on AI for film directors. Something that feels like a gallery catalogue — lots of whitespace, editorial serif for the headings. Hero at the top with room for a trailer video, three tiers of pricing, a section with quotes from students, an FAQ at the bottom. The whole thing should feel unhurried.
Claude renders. You look at it. You iterate.
The iteration room — sketches scattered, one live canvas on a tablet, the work moving between hand and screen. Claude Design sits in the gap between “I know roughly what I want” and “I can show someone.”
There are two ways to iterate. For big structural changes — “move the pricing above the testimonials, add a section about who it’s for” — you chat. For small component-level changes — “this heading is too big, this button should be cream not black” — you comment inline on the element itself, the way you’d leave a note in Figma. Mixing the two is what keeps the work moving without losing state.
There are also sliders. Anthropic calls them adjustment knobs. They’re live controls for spacing, typography density, colour temperature, layout grid — generated on the fly for whatever you’re working on. You nudge them, the canvas updates.
Where it goes when you’re done
Export paths matter. Claude Design lets you leave through four doors:
- Canva. Direct integration — Melanie Perkins is quoted in Anthropic’s announcement. Your design lands in Canva with your Brand Kit already applied.
- PDF or PPTX. Editable, not a print-to-PDF flatten. Someone can actually edit the slides after they land.
- Standalone HTML. Download and host anywhere.
- A Claude Code handoff bundle. This is the interesting one. When the design is ready to actually ship as a real app or site, Claude Design packages everything — the components, the design system, the structure — into a bundle you hand to Claude Code. Claude Code then builds it for real.
You can also share a read-only URL inside your org, or grant comment-only or edit access. Organisation-scoped by default, which is the right default.
The honest caveat: token burn
Here is where the marketing stops and the reality starts.
Claude Design runs on Opus 4.7. It produces interactive HTML with vision and iteration. That is the most expensive kind of work a model can do per turn. One PCWorld reviewer burned through roughly 80% of their weekly Pro allowance in thirty minutes of design work, then got locked out entirely when an undo failure forced a rebuild.
The Pro plan is not the right plan for daily Claude Design work. It’s the right plan for occasional, deliberate use. If you’re going to lean into it — make it part of a workflow, iterate all week — you want Max.
A second caveat, which matters: the undo button can be dangerous. There are reports of sessions where a single undo caused Claude to report it had wiped the files and needed to rebuild everything from scratch. Save frequently. Export early versions even if you think you’ll keep iterating.
And the last: it’s a research preview. Known teething problems at launch — inline comments occasionally disappear before the model reads them, the compact layout can throw save errors, large repositories cause browser lag, GitHub integration was patchy on day one. Live with it.
Sessions: parallel universes in Claude Code
The Claude Code desktop rebuild that dropped on Tuesday had a lot in it, but the headline feature is Sessions.
A session is an independent instance of Claude Code, with its own context and its own repository view. The redesign makes the sidebar the centre of the app. Every session you’ve ever started — running, waiting on input, archived — lives there, filtered by status, grouped by project.
What this unlocks is a different way of working. Before, if you wanted Claude to do two things at once, you opened two windows and hoped they didn’t step on each other. Now you have one app, many lanes, and each lane has its own memory.
A real use case, from how I’ve been working this week:
- Lane 1: writing this article in one repo, Claude pulling research and drafting as I go
- Lane 2: a second session in the art-and-algorithms repo fixing a small Stripe webhook bug
- Lane 3: running a shoot script in yet another folder, generating reference images
Three things moving in parallel, none of them confused about what the others are doing. Each one has its own context window. Each one has its own tools available. I glance at the sidebar, I see where each one is, I jump in where I’m needed.
Alongside Sessions, the redesign brought a handful of other useful things:
- Side Chat Shortcut (Cmd + ;). Branch a quick question off a running task without polluting the main thread’s context. “Hey quick — what does this error mean?” opens in a scratch chat. The main thread keeps its focus.
- Integrated terminal and file editor. Tests, builds, spot edits without leaving the window.
- Three view modes. Verbose shows every tool call Claude makes, Normal is the default, Summary hides everything but the decisions. Pick what you want to see.
- Drag-and-drop layout. Arrange the terminal, preview, diff viewer, and chat in whatever grid matches how you work.
The cost multiplier
Sessions are genuinely powerful. They are also genuinely expensive.
Running four sessions in parallel at 100K tokens each is 400K tokens of work, not 100K. This is obvious when you say it out loud and invisible when you’re in the middle of it. Hands-on reviews found Pro users exhausted their weekly quota in ninety minutes of enthusiastic four-lane work.
The right default is probably one or two active sessions, occasionally three. Four or more is a specific situation — a big refactor where you want tests running in parallel with the implementation — not a normal workday.
Routines: Claude that works while you’re away
Routines shipped the day before Sessions. Less visible, arguably more important.
A routine is a scheduled Claude Code job. You bundle a prompt, a repo, and any connectors — GitHub, MCP servers (the open protocol that lets Claude talk to Stripe, Notion, Supabase, and anything else with an MCP integration), whatever you’ve wired up — into a single saved configuration. Then you set it to run on a schedule, or fire from an API call, or trigger off a GitHub event — a new pull request, a merged branch, an issue being opened.
Crucially, routines run on Anthropic’s cloud, not locally. Your laptop doesn’t need to be awake. You can close the app, go for a walk, come back to the work done.
The plan limits matter (9to5Mac’s breakdown):
- Pro: 5 routines per day
- Max: 15 routines per day
- Team and Enterprise: 25 routines per day
A routine running at night. The room is empty, the chair is pushed back, the monitor is still working.
The creative examples are where this gets interesting. Routines isn’t just for engineering teams running nightly test suites. A few things I’m either doing or planning to do:
- Every Monday at 9am, look at what I shipped to LinkedIn last week, pull the analytics, and draft three follow-up posts in my Obsidian vault.
- When a new article merges to main on this site, run the SEO link audit and post a summary of any gaps to my daily log.
- Every Friday evening, look at this week’s shoot outputs, grade them, and draft the next week’s shoot plan based on what worked.
The pattern is the same: a small, boring, repeatable task that used to eat twenty minutes of the day. Bundle it, schedule it, forget it.
The difference from a cron job
You might reasonably ask: isn’t this just cron with a language model attached? Sort of. The difference is that a routine has judgement. It reads the inputs, thinks about them, makes a call, writes the output. A cron job runs the same script with the same parameters whether or not the situation warrants it. A routine can decide “actually, there’s nothing worth posting about this week — skip.”
That decision-making is what makes this feel different. It’s also what makes it cost tokens. Every routine run is a Claude call, and Claude calls are not free. Use them for work where the judgement is the point.
Opus 4.7: the quiet engine under everything
The model that powers the above three products went generally available on Thursday, between the desktop rebuild and Claude Design. It kept the same pricing as Opus 4.6 — $5 per million input tokens, $25 per million output tokens — which is quietly the biggest news of the week. Every announcement that followed costs the same to use as the previous model did.
Three things matter for creative work:
- Better at long runs. Holds its train of thought across hours of agentic work, notices its own logical faults mid-plan, makes fewer tool errors. This is the reason Claude Design can iterate on a design for thirty minutes without losing the plot — and the reason parallel Sessions don’t drift.
- Better vision. Images now process at up to 2,576 pixels on the long edge — roughly three times the megapixel count of the previous version. Long screenshots, dense diagrams, scans of handwritten notes all come back more accurately.
- Better coding. Resolves three times as many production tasks as Opus 4.6 on Rakuten-SWE-Bench, 70% on CursorBench versus 58% before. Numbers aside, what this feels like day-to-day is fewer “wait, you broke that other thing” moments when Claude edits code.
There are also a handful of developer-focused additions — a new xhigh effort level for hard reasoning, Task Budgets in public beta (a ceiling on how many tokens a long-running task can consume), the /ultrareview command for deep code review, and Auto Mode extended to Max users. None of these change the shape of the product. They just give people who live in Claude Code more knobs.
The engine got quieter. Which is really the point.
The courses page: where to start if this is new
If any of the above felt foreign, Anthropic has a free teaching layer at claude.com/resources/courses. Sixteen courses, video lessons, certificate on completion.
For a vibe coder picking up Claude for the first time, the sensible path is Claude 101 → Claude Code 101 → Claude Code in Action. Three hours total, the first two an hour each. You’ll come out with enough working knowledge to use everything this week shipped.
If you want to build deeper — wiring Claude to Stripe or Supabase or Notion through the Model Context Protocol — Introduction to MCP and its advanced follow-up are the route in.
How they compose, at least on my desk
Anthropic hasn’t published a roadmap tying these launches together — the one place the composition is explicit is the handoff bundle from Design to Code. Everything else I’m about to describe is how they’ve been sitting on my desk this week, not a claim about design intent.
Here’s the loop I’ve ended up running.
I use Claude Design to think visually — sketch a landing page, prototype a feature, mock up a pitch. I work on the canvas until the shape is right. Then I hand it to Claude Code via the handoff bundle. Claude Code builds the real thing — the actual site, the actual components, matched to my design system.
I run that work across parallel Sessions when the build is big — one session on the frontend, one on the backend, one writing tests — glancing at the sidebar to see where each one is. Then I set a Routine to keep the boring parts running — the weekly SEO audit, the Monday content draft, the Friday retrospective. While I’m asleep, while I’m on holiday, while I’m doing other work.
Opus 4.7 is under all of it, slightly better at each step than Opus 4.6 was last week.
Maybe the pieces were designed to fit together like this. Maybe I’m reading coherence into four shipping events that happened to land in the same week. Either way, the loop works.
What’s still missing
Three absences worth naming, on top of the caveats already in each section:
- No image generation in Claude Design. If your design requires photography, you’re leaving and coming back. That’s a real constraint for anything marketing-led.
- Claude Design reads your codebase, not your Figma. If your design system lives only in Figma, the read is thinner. If it lives in code — a Tailwind config, a tokens file, component files — the read is accurate.
- None of this is GA together. Claude Design is a research preview. Routines are new enough to have daily caps. Sessions will bite on token cost the moment you stop counting. Treat the whole stack as experimental for another month or two.
Where to go next
If you’re starting from zero with any of this, a non-developer’s first hour with Claude Code is the right place to begin — it was written before this week’s changes, and the core ideas still hold. Pair it with Claude 101 if you want a guided run through.
If you already use Claude Code and want to push further, autopilot hooks and parallel agent teams both foreshadowed Sessions and Routines — the patterns those articles describe are now first-class features.
And if you’re going to try Claude Design this weekend, start small. One landing page. One pitch deck. Watch how it reads your design system. Watch how the handoff bundle lands in Claude Code. Notice what it can’t do. Come back next week with half a sense of it.
The pace isn’t going to slow down. Better to know the shape now.