Back to Blog
claude-code weekly-update opus-4-7 routines claude-design glasswing community

What's Up Claude Code: Week of April 12th

Olivier Legris ·

TL;DR

The biggest Claude Code week since Opus 4.6. Opus 4.7 went GA on April 16 with a new xhigh effort level (now Claude Code’s default) and 3× higher-resolution vision — reception is split. Routines turn Claude Code into a scheduled, webhook-triggered cloud workflow — the conceptual leap of the week. Claude Design shipped as a research preview, taking direct aim at Figma. The CLI release train hit v2.1.114, including a move to native per-platform binaries. And Project Glasswing / Mythos Preview published results: thousands of zero-days across every major OS, including a 17-year-old FreeBSD NFS RCE.

1. Opus 4.7 — And Why xhigh Is the New Default

Opus 4.7 went GA on April 16, priced identically to 4.6 ($5/$25 per M tokens). Available on Claude products, the API, Bedrock, Vertex, and Microsoft Foundry.

The three practical changes:

  • New xhigh effort level, sitting between high and max. Claude Code now defaults to xhigh on all plans. If you’ve been manually bumping effort, stop — you’re already there.
  • High-res vision: max image size jumps from 1568px / 1.15MP to 2576px / 3.75MP — roughly 3×. Screenshots of dense dashboards or long-form mockups finally work without pre-processing.
  • Task budgets: a rough token target spanning the full agentic loop (thinking + tool calls + tool results + output), not just output tokens.

A new /ultrareview slash command lands alongside for thorough code review passes. It was parallelized in v2.1.113 — use it on PRs before merging, not on every file.

Reception is split. Benchmarks show gains (a community blind-eval with three judges put 4.7 at 69% wins vs 4.6), but early users are flagging real regressions: a new tokenizer that consumes ~1.35× more tokens for the same prompt, MineBench creative-task drops vs 4.6, and qualitative “conversational coherence” complaints. Combined with the cache TTL change on April 2 (1h → 5m default), the “4.7 is more expensive” narrative is gaining ground fast. Watch the next two weeks for Anthropic’s response — if it doesn’t come, the sentiment will ossify.

2. Routines — Claude Code Leaves the Laptop

Announced April 14: Routines in Claude Code, research preview.

A Routine is a prompt + repo + connectors, configured once, then run automatically by one of three triggers:

  • Schedule (hourly / nightly / weekly)
  • API call (each Routine gets an HTTP endpoint + bearer token)
  • GitHub webhook (repo events)

Runs on Anthropic’s web infrastructure. Your laptop doesn’t need to be open. Quotas: Pro 5/day, Max 15/day, Team/Enterprise 25/day.

This is the conceptual shift — Claude Code crossing from “tool you sit next to” to “cloud workflow you configure once.” Early patterns emerging in the first week: backlog triage on cron, docs-drift scans on push, deploy verification on release, alert triage on webhook, dependency porting across repo sets.

Launched alongside a full Claude Code desktop app redesign for Mac and Windows. The desktop app is now the Routines control surface as well as the local session UI.

3. Claude Design — A New Surface, Aimed at Figma

Claude Design shipped April 17 as a research preview — Opus 4.7-powered generation for prototypes, slides, one-pagers, and mockups. It reads your codebase and design files to apply your company’s design system consistently, then exports to PDF, URL, PPTX, or pushes to Canva.

Available to Pro, Max, Team, and Enterprise. Anthropic’s CPO resigned from Figma’s board three days before launch. Figma closed -4.26% on the announcement. Adobe, Wix, and GoDaddy also traded down.

The interesting piece for Claude Code users: the design system awareness runs off the same codebase indexing, which means the PM-side and dev-side output start from the same source of truth. That’s new, and it’s the feature worth watching.

4. The Release Train: v2.1.110 → v2.1.114

Six releases in five days. Organizers fielding “what changed” questions, the cheat sheet:

  • v2.1.110 (Apr 15) — /tui command, fullscreen flicker-free rendering, Ctrl+O verbose transcript toggle, push notifications for Remote Control, autoScrollEnabled config.
  • v2.1.111 (Apr 16) — Major. Opus 4.7 xhigh support, interactive /effort slider, “Auto (match terminal)” theme, /less-permission-prompts skill, /ultrareview, Windows PowerShell tool rollout, read-only bash no longer prompts, improved /setup-vertex and /setup-bedrock.
  • v2.1.112 (Apr 16) — Hotfix for “claude-opus-4-7 temporarily unavailable” in auto mode.
  • v2.1.113 (Apr 17) — Major. CLI now spawns a native per-platform binary instead of a bundled JS runtime. Quiet but significant: faster cold start, smaller install. New sandbox.network.deniedDomains setting. Fullscreen Shift+↑/↓ viewport scroll, Ctrl+A / Ctrl+E logical-line nav, Windows Ctrl+Backspace word delete, /loop Esc cancels wakeups, /extra-usage works from Remote Control, parallelized /ultrareview. Security: macOS dangerous-path handling, Bash deny-rule hardening.
  • v2.1.114 (Apr 18) — Crash fix in the permission dialog when an agent-teams teammate requested tool permission.

If you’re still on the v2.1.104 line from two weeks ago, the native-binary move alone is worth the upgrade.

5. Project Glasswing & the Mythos Preview

Mythos Preview — a frontier model gated to defensive cybersecurity researchers only — published results this week. Thousands of zero-days identified across every major OS and browser, including triage of CVE-2026-4747, a 17-year-old RCE in FreeBSD NFS.

Partners on Project Glasswing: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia. Anthropic committed $100M in model credits, plus $2.5M to Alpha-Omega / OpenSSF and $1.5M to the Apache Software Foundation.

Worth noting the pushback: Tom’s Hardware reported that the “thousands of zero-days” figure traces to 198 manually-reviewed cases, with the rest extrapolated. Even discounted, it remains the first public-facing datapoint on directed offensive-security use of a frontier model — and the direction is defense. Expect this to reshape the CVE backlog conversation for the rest of the year. (Schneier’s take.)

6. Deprecation Notice — Haiku 3 Retires April 19

claude-3-haiku-20240307 goes offline tomorrow. If you’re still pinning it in production code, migration path is Haiku 4.5 (claude-haiku-4-5-20251001). Check your batch jobs, your MCP servers, and any crusty CI helper scripts before Monday.

7. Community Spotlight

Six projects worth knowing from this week, themed around the week’s two dominant anxieties: where are my tokens going and what does Claude Design mean for the rest of my stack.

  • Claude Code token attribution TUI — Reads session transcripts from ~/.claude/projects/ and classifies every turn across 13 categories (debugging, brainstorming, refactor, etc.) via deterministic tool-usage patterns — no LLM calls. Author reports $200+/day spend visibility. The ccusage-for-why tool the community has been asking for.
  • /extract-design — Claude Code plugin that pulls a full design system from any URL (colors, fonts, spacing, shadows, component patterns) into Claude-structured markdown + a Tailwind config. Shipped the same week as Claude Design; the counter-move of “bring any site’s design system into Claude Code” is a useful complement.
  • Nelson (250+⭐) — Multi-agent coordination skill using Royal Navy command structure (admiral → captains → ships → crew). Risk tiers gate human approval, damage control handles stuck agents and context exhaustion. Shipped v2.0 with persistent agent memory this week.
  • evo (autoresearch for code) — /evo:discover picks an optimization target in a repo (benchmark, pass rate, latency); /evo:optimize spawns parallel agents to improve it. Karpathy’s autoresearch loop applied to codebases instead of ML training. An early, useful scaffolding for Routine-driven optimization work.
  • AgentDiff — Line-level attribution across multi-agent codebases. agentdiff list shows which agent (Claude Code / Cursor / Codex), which prompt, and the intent behind every line. Compliance-oriented — aimed at enterprise adoption where audit trails are mandatory. The natural pair to any Routines production rollout.
  • ResolveMCP — MCP server connecting Claude to DaVinci Resolve Studio. Timeline editing, color nodes, Fusion comps, Neural Engine features (Magic Mask, Smart Reframe) — all via natural language. Good signal on the “MCP as universal creative-tool adapter” trend.
  • Boris Cherny’s 6 Claude Code tips for Opus 4.7 — community-curated from Boris’s comments in the GitHub caching thread. Actionable: be selective about skills/agents per project; start fresh conversations instead of leaving sessions open past cache expiry; audit context before typing.
  • Blind-eval Opus 4.7 vs 4.6 — 100 questions, three-judge panel (GPT-5.4, Gemini 3.1 Pro, DeepSeek V3.2). 4.7 wins 69% aggregate; notably, DeepSeek disagrees (41%) — a clean illustration of eval-judge bias in cross-model benchmarks.
  • Context audit: 30K tokens gone before typing — Author found ~170 agents/skills auto-loading, ~10 actually used. Reinforces Boris’s guidance. The “skills-as-retrieval” thesis keeps getting louder.
  • Cache TTL regression GitHub issue — “Cache TTL silently regressed from 1h to 5m around early March 2026.” Cache write rate is 12.5× cache read rate. This is the invisible variable behind most “Opus 4.7 is expensive” posts.

Patterns worth watching: the visibility layer is filling in. Token attribution, agent attribution, and context audit tools all shipped this week — a direct response to a month of “where did my budget go” frustration. Expect Anthropic to merge some of this into Claude Code native within 60 days.

8. Community Events

181 events, 215 leaders. A fuller week than last, spanning four continents.

Coming next week (Apr 20–26):

New cities this week: Aix-en-Provence 🇫🇷, Aarhus 🇩🇰

Host your own on Luma →

What We’re Watching

  1. Routines adoption curve — Cron-replacement is the obvious first use case. The interesting question is whether people trust Routines for write operations (PRs, deploys, emails) or keep them read-only for the first quarter.
  2. xhigh as a silent quality bump vs. silent cost bump — Most users won’t realize their defaults changed. Expect a confusing mix of “Claude got smarter” and “Claude got expensive” threads — both are true, and the tokenizer change makes them the same story.
  3. Claude Design vs. the codebase-of-truth story — If design generation actually stays consistent with the repo’s design system across weeks, this is the first product that makes “design/dev single source of truth” real.
  4. Native binary CLI — The v2.1.113 shift from bundled-JS to per-platform native binaries is a long-term performance and security story. Watch install size, cold-start latency, and how MCP subprocess behavior changes.
  5. Glasswing follow-through — $100M in credits to defenders is a commitment, but the practical question is how quickly the CVE backlog actually moves. The FreeBSD NFS fix is the first datapoint. Look for the second one within 30 days.
  6. The visibility layer — Token attribution, agent attribution, and context audit tools all shipped this week from the community. Expect Anthropic to absorb some of this natively within 60 days; the gap between “I’m paying” and “I know why I’m paying” has become the sub’s #1 frustration.