Anthropic released Claude Opus 4.7 today (April 16, 2026). It's the successor to Opus 4.6, and it came bundled with a pretty significant Claude Code desktop redesign. I've been poking around with it all morning, so here's the rundown of what's new and what's actually worth caring about.

Opus 4.7 is better at coding. That's the headline. On SWE-bench Pro, it scores 64.3% — up from 53.4% on Opus 4.6. For context, GPT-5.4 sits at 57.7% and Gemini 3.1 Pro at 54.2%. On SWE-bench Verified, Opus 4.7 hits 87.6%, compared to 80.8% on its predecessor.

But the raw benchmarks aren't the interesting part. What I've noticed is how much better it handles long-running agentic tasks. It plans more carefully before jumping into code, catches its own mistakes more often, and follows instructions with almost annoying precision. If your prompts were loose and it worked fine with 4.6 — you might need to tighten them up. Opus 4.7 interprets directions literally.

On the Rakuten-SWE-Bench (which tests against production-scale codebases), it resolves 3x more tasks than Opus 4.6, with clear gains in both code quality and test quality. On CursorBench, it clears 70% versus 58% before.

Opus 4.7 now processes images at up to 2,576 pixels on the long edge — roughly 3.75 megapixels. That's more than triple what previous Claude models could handle. This matters if you're building computer-use agents, doing diagram analysis, or anything where you need the model to actually read what's on a screen or in a technical drawing. Chemical structure diagrams, complex architecture charts — stuff that used to be too blurry for the model to parse.

There's a new tokenizer. It processes text more efficiently, but the same input can map to 1.0–1.35x more tokens depending on content type. So your token bills might tick up slightly for the same work. Not a dealbreaker, but worth knowing — especially if you're running high volumes through the API.

Anthropic added a new effort level called xhigh that sits between high and max. Claude Code defaults to xhigh on Opus 4.7 for all plans. You get more thinking depth than high without going full max (which has no token spending constraint at all). It's a practical middle ground — better reasoning without burning your budget on every request.

Honestly, the Claude Code updates might matter more day-to-day than the model upgrade itself. Anthropic shipped a full desktop redesign on April 14, two days before the model dropped.

The desktop app now has a sidebar for managing multiple sessions. You can run several Claude instances side by side in the same window. Start work across different repos, switch between them as results come in, filter by status or project. Sessions auto-archive when PRs merge or close.

Sounds minor. Changes your workflow. Before, you'd have multiple terminal tabs and lose track of what's running where. Now it's all in one place with proper state management.

Hit ⌘+; (or Ctrl+; on Linux/Windows) and you get a side chat that pulls context from your main thread but doesn't write back to it. Perfect for quick questions about the codebase without derailing your current task. Small feature, big quality-of-life win.

This is the headliner. Routines let you package up a prompt, connect it to your repos and connectors, and set it to run automatically. Three types of triggers:

The important part: routines run on Anthropic's infrastructure, not your machine. Your laptop can be off and your routines still execute. Think automated code reviews on every PR, scheduled dependency audits, nightly test runs with AI-driven failure analysis.

Plan limits: Pro gets 5 routines/day, Max gets 15, Team and Enterprise get 25.

There's a new /ultrareview command that kicks off a multi-agent review session on your PRs. It's meant for large or critical changesets — 500+ lines — and takes 10-20 minutes to produce verified results. The regular /review takes 3-4 minutes and catches common issues. /ultrareview goes deeper. Pro and Max users get three free ultrareviews.

Integrated terminal, rebuilt diff viewer (handles large changesets better), in-app file editor, HTML/PDF preview pane. Three view modes — Verbose, Normal, Summary — so you can control how much output you see. SSH support is now available on Mac and Linux. Drag-and-drop layout for arranging panes however you want.

Unchanged. $5 per million input tokens, $25 per million output tokens. Available through the API (model ID: claude-opus-4-7), Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry.

One thing to watch: the new tokenizer means the same text can produce slightly more tokens, so your effective cost per request edges up a bit even though the per-token price didn't change.

Opus 4.7 is a solid iteration, not a paradigm shift. The coding improvements are real and measurable — I've already seen it handle multi-file refactors that 4.6 would get lost in. Vision goes from "kinda useful" to "actually usable for real work." The stricter instruction following is a net positive even if it means some prompt tuning on your end.

But the Claude Code updates are what I'm most interested in. Routines turn Claude Code from something you sit and watch into something that works while you sleep. The multi-session sidebar removes real friction. /ultrareview fills a gap that regular review commands couldn't touch.

If you're on Claude Code already: update, try the sidebar, and set up at least one routine. If you're using the API: swap to claude-opus-4-7 and test your existing prompts — the stricter instruction following means some of them will behave differently. And if you're running vision workflows, the 3x resolution bump is worth re-evaluating what's possible.