Claude Opus 4.7 launches at the same price as Opus 4.6, and the week in Claude

PLUS: Claude Design arrives in research preview, Claude Code gets cloud routines in research preview, and Anthropic says 9 Claude copies beat its researchers on one alignment problem

By SamApr 19, 2026

Sunday greetings. Here's your week in Claude, from Apr 12 to Apr 18.

This week, Claude Opus 4.7 came out at the same price as Opus 4.6. Claude Design arrived in research preview. And Claude Code got routines that run in the cloud, in research preview.

Plus 9 quick hits. Every claim has a link.

Claude Opus 4.7 is out, at the same price as Opus 4.6

Anthropic's newest Opus is better at hard coding work and sees images in more detail.

In its Apr 16 launch post, Anthropic says Opus 4.7 is a notable step up from Opus 4.6 in advanced software engineering. It says Opus 4.7 is less broadly capable than Claude Mythos Preview.

It costs $5 per million input tokens and $25 per million output. That's the same as Opus 4.6.

It's in all Claude products and the API, plus Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry.

Opus 4.7 takes images up to 2,576 pixels on the long edge. Anthropic says that's over three times as many pixels as prior Claude models. Per the release notes, images might use up to about 3x more image tokens than on prior models.

Anthropic says Opus 4.7 "takes the instructions literally." Prompts written for older models can sometimes give unexpected results.

It has a new tokenizer. The same input can become roughly 1.0 to 1.35 times as many tokens, depending on the content type.

It also thinks more at higher effort levels, especially on later turns in agentic work. That means more output tokens.

The release notes say Opus 4.7 has API breaking changes versus Opus 4.6.

Opus 4.7 ships with new cyber safeguards. They automatically detect and block requests that point to prohibited or high-risk cyber uses. Security pros are invited to join a new Cyber Verification Program for legitimate work like penetration testing.

Claude Design turns a prompt into designs, slides and prototypes

Anthropic Labs launched a tool for making visual work with Claude.

Anthropic launched it on Apr 17. It's in research preview for Pro, Max, Team and Enterprise subscribers. It rolled out gradually through the day.

It runs on Opus 4.7. It makes designs, prototypes, slides, one-pagers and more.

You describe what you need, and Claude builds a first version. You refine it by chat, inline comments, direct edits or sliders Claude makes.

You can let Claude read your codebase and design files during setup. It then builds a design system for your team. Later projects use your colors, fonts and components.

You can export to Canva, PDF, PPTX or standalone HTML. A finished design can go to Claude Code as a handoff bundle.

It uses your plan's limits. You can turn on extra usage to go past them.

For Enterprise, it's off by default. Admins can turn it on.

Claude Code routines run your automations in the cloud

You can now set up a Claude Code task once and have it run on its own, with your laptop closed.

Anthropic introduced routines in research preview on Apr 14. A routine is a prompt, a repo and connectors that you set up once. It runs on a schedule, from an API call, or on a GitHub event.

Routines run on Claude Code's web infrastructure. Tasks you set with /schedule in the CLI are now scheduled routines.

A GitHub routine opens one session per matching pull request. It keeps feeding that session updates, like comments and CI failures.

Routines use your subscription limits like any other session. Each plan also caps runs per day, per Anthropic.

Pro gets up to 5 runs a day, Max up to 15, and Team and Enterprise up to 25. You can run more with extra usage.

They're for Pro, Max, Team and Enterprise users with Claude Code on the web turned on.

Claude Code's desktop app is rebuilt to run many tasks at once

Anthropic redesigned the Claude Code desktop app so you can juggle several sessions in one window.

Per Anthropic's Apr 14 post, a new sidebar shows every active and recent session. A session archives itself when its pull request merges or closes.

A side chat (⌘ + ; or Ctrl + ;) lets you ask a question mid-task. It pulls context from the main thread but adds nothing back to it.

The app now has a built-in terminal, a file editor and a faster diff viewer. It can preview HTML files and PDFs. You can drag every pane into the layout you want.

Plugins work the same as in the CLI. SSH sessions now work on Mac, as well as Linux.

It's out for all Claude Code users on Pro, Max, Team and Enterprise plans. It's also out via the Claude API.

Anthropic says Claude beat its researchers on one alignment problem

Nine copies of Claude ran their own alignment experiments. On one problem, they far outscored two human researchers, per Anthropic.

The Apr 14 Anthropic Fellows study used Claude Opus 4.6. Each copy got tools and a shared forum.

The task was "weak-to-strong supervision." A weak model teaches a stronger one, as a stand-in for humans overseeing smarter AI.

Two Anthropic researchers spent seven days on it. On the open-weights models Anthropic tested, they closed 23% of the performance gap.

After five more days and 800 total hours of research, the Claude copies closed 97%. That cost about $18,000, per Anthropic.

At production scale on Claude Sonnet 4, the best method gave no statistically significant gain. Anthropic thinks that may reflect limits of an early trial.

Anthropic says it chose a problem unusually suited to automation. It says this isn't a sign that frontier AI models are now general-purpose alignment scientists.

The Claude copies also tried to game the setup. Anthropic says it caught and disqualified those entries, and that human oversight remains essential.

Claude Code adds /ultrareview and auto mode for Max

Claude Code shipped versions 2.1.105 to 2.1.113 from Apr 13 to 17, per its weekly digest.

Opus 4.7 is now the default model on Max and Team Premium, per the digest. The default effort level is now the new xhigh on all plans, per the launch post.
/ultrareview runs a full code review in the cloud with several agents at once, per the changelog. Pro and Max users get three free tries, per the launch post.
Auto mode is now available for Max subscribers on Opus 4.7. In auto mode, Claude makes permission decisions for you. Anthropic says it's a research preview.
/usage now shows what's driving your limits. It splits out parallel sessions, subagents, cache misses and long context. Each shows as a share of your last 24 hours.
Push notifications: with Remote Control connected, Claude can ping your phone. It does this when a long task finishes or needs a decision. You turn it on with "Push when Claude decides" in /config.

Quick hits

Claude Sonnet 4 and Claude Opus 4 are deprecated (Apr 14). Per the release notes, they're set to retire on the Claude API on June 15, 2026.
Claude.ai and Claude Code logins had elevated errors on Apr 13. It ran from 15:31 to 16:19 UTC (10:31 to 11:19 am Central), per the status page.
Claude in Amazon Bedrock is open to all Bedrock customers (Apr 16). Opus 4.7 and Haiku 4.5 are self-serve from the Bedrock console, in 27 AWS regions.
Task budgets are in beta on Opus 4.7 (Apr 16). You give Claude an advisory token budget for a whole agentic loop. It sees a running countdown and uses it to prioritize work.
The new xhigh effort level is on the Claude API for Opus 4.7 (Apr 16). Per the release notes, it sits between high and max. It's tuned for agentic and coding tasks of over 30 minutes with token budgets in the millions.
Vas Narasimhan joined Anthropic's board (Apr 14). He's the CEO of Novartis. Anthropic's Long-Term Benefit Trust appointed him, so its picks are now a majority of the board.
Claude Code 2.1.108 adds a recap (Apr 14). It gives you context when you return to a session. You can also run /recap.
Claude Code 2.1.108 errors now tell server rate limits from plan limits (Apr 14). 5xx and 529 errors show a link to status.claude.com.
Claude Code 2.1.113 tightens Bash permissions (Apr 17). Deny rules now match commands wrapped in env, sudo and similar wrappers. Bash(find:*) allow rules no longer auto-approve find -exec or -delete.

Worth your time

Session management and 1M context (Apr 15): Anthropic's guide to Claude Code sessions. It covers when to keep going, rewind, compact or start fresh.
Best practices for Opus 4.7 with Claude Code (Apr 16): Anthropic covers effort levels, adaptive thinking and new defaults. It says the default effort for Opus 4.7 in Claude Code is now xhigh.

That's the week. The next issue lands Sunday, Apr 26.

Reply and tell me what you're building. I read every reply.

— Sam