Claude Report — 2026-09-27
- Opus 5.5 gets a coding-focused update: 20% cheaper tokens, 60% cheaper cache reads, ~30% faster output (Sep 24).
- Claude Tag adds personal connectors inside shared Slack channels, rolling out on Team plans (Sep 24).
- Anthropic research: Claude computes a nine-loop physics amplitude for ~$1-2K total compute, independently validated by a SLAC physicist (Sep 25).
- <a href="https://github.com/nicobailon/pi-subagents">nicobailon/pi-subagents</a> ships two releases (v0.72.0/v0.72.1) after the most active four-day sprint of any tracked repo.
- New community tools ship for chess-game postmortems, bash-command guardrails, and cross-agent MCP server toggling.
🚀 Anthropic Official
Claude Opus 5.5: built for coding sessions that use more context (Sep 24)
Cuts Opus 5.5 pricing ~40% for typical workloads: input/output tokens down 20%, cached-token reads down 60%. Claude Code changes cut cache misses by over 50% by preventing accidental cache breaks on login refreshes and mid-conversation instructions, letting effort levels change without resetting cache, extending one-hour cache lifetime to API users, and letting forked subagents inherit the parent's cache instead of repaying for identical context. Opus 5.5 also needs fewer turns and generates output ~30% faster than Opus 5 on long, context-heavy dev sessions.
Claude Tag supports personal connectors in channels (Sep 24)
Lets individual Slack users invoke their own personal connectors (calendar, drive, CRM accounts, staging deploys) inside a shared channel, instead of only admin-configured shared tools; only the connecting user can trigger their own connectors. Ships with a review-before-post mode or auto-post unless sensitive content is detected, plus admin controls for shared-identity, personal-only, or per-tool governance. Rolling out now on Team plans, Enterprise to follow.
Yes, Claude can do nine loops (Sep 25)
Claude Science running Fable 5.1 computes a nine-loop six-particle (hexagon) amplitude in planar N=4 super Yang-Mills theory for roughly $1,000-2,000 total compute (the bootstrap piece alone cost ~$100, equivalent to 96 CPUs for a week), using Python/SymPy and succeeding on the first attempt with essentially no scientific oversight beyond "keep going." SLAC physicist Lance Dixon independently verified the result, and a separate team at the Chinese Academy of Sciences (Song He's group) obtained convergent results using AI assistance.
Using Claude Code: Spending your effort (Sep 25)
Anthropic's Thariq Shihipar publishes a Claude Code engineering post on the effort parameter, showing Fable 5.1 improving from 1/5 to 5/5 success on a security-sanitizer task as effort rises, with Terminal-Bench 3.0 data backing the analysis. Recommends low effort for quick iteration/brainstorming (~2 min), medium for standard feature work, high for bug fixes and edge cases, and max effort for autonomous problem-solving (up to 79 minutes observed) — with the biggest gains on security, hardware, and ML edge cases rather than routine tasks.
🔌 Claude Code Plugins
Plugin tracking skipped this run — official marketplace install-count cache unavailable.
🛠️ Skills
addyosmani/agent-skills
Bumps plugin manifests to v0.6.11 and merges hardening fixes (Sep 24-26): guards artifact/spec-plan-todo paths, strips negated trigger clauses fully instead of partially, enforces skill-anatomy layout rules, and validates manifest versions against the root plugin.json.
anthropics/skills
Updates the claude-api skill to link the refusal-billing change to the docs (Sep 24), the repo's only commit in the window.
Unchanged in window: cloudflare/security-audit-skill (last commit Sep 14), K-Dense-AI/scientific-agent-skills (last commit Sep 21).
🤖 Agents & Subagents
nicobailon/pi-subagents
Ships v0.72.0 and v0.72.1 (Sep 27) fixing workflow result/artifact handling and unknown-agent validation, capping a four-day sprint covering lazy-loading performance, provider-prefixed model ID resolution, capability-ceiling enforcement, and async-child cost tracking (Sep 24-26).
Unchanged in window: VoltAgent/awesome-claude-code-subagents (last commit Sep 21), anthropics/commerce-agents (last commit Aug 31).
🔗 MCPs & Integrations
DeusData/codebase-memory-mcp
Merges a JAX-RS relative-path fix and a sanitized build-config PR (Sep 27), following Windows worker-recovery fixes (Sep 25) and service-pattern-boundary work (Sep 24).
upstash/context7
Documents final Search API pricing (Sep 25) and adds a Docs7 custom-domains page (Sep 24), continuing its documentation push.
ChromeDevTools/chrome-devtools-mcp
Fixes --blockedUrlPattern/--allowedUrlPattern rejecting hostname regexp groups and empty filePath defaulting to an empty string instead of undefined (Sep 25).
oraios/serena
Lands one small fix correcting a duplicated "IS" string in the ignored-path CLI check output (Sep 24).
Unchanged in window: github/github-mcp-server (last commit Sep 16), modelcontextprotocol/servers (last commit Sep 22).
💡 Community — Workflows & Ideas
brumar/chess-postmortem-skills
Ships a Claude Code skill set that turns a chess game into a postmortem: Stockfish-backed mistake analysis via investigator subagents, think-aloud audio aligned to moves, annotated PGNs, and narrated analysis videos (Sep 26, 73 pts on HN).
SeanMcMillan/quiet-guardrails
Adds a Claude Code PreToolUse hook that auto-corrects inefficient or risky bash patterns deterministically instead of prompting for approval each time, while still forcing confirmation on genuinely risky ops like git add (Sep 25).
lava/withmcp
Releases a Rust CLI for toggling MCP servers system-wide or per-project across Claude, Codex, and Pi, with saved profiles and path-based enable/disable rules (Sep 26).
davila7/claude-code-templates
Adds new mods: a chess mod where a Jev-configured key plays instead of Claude, a jev-vercel-sandbox mod that runs dangerous Bash commands inside a Vercel Sandbox, and a catalog fix that stops publishing partial download counts (Sep 26).
Cockpit
Ships a free local macOS dashboard that sits above Claude Code/Codex CLIs, showing unified session cards, an "attention" view for sessions awaiting input, and in-app reply/permission approval (Sep 25).
📰 Quick Mentions
- tenjin-agent — Jev-based x402 micropayment tool router that auto-pays for external tools (Context7, Exa, Firecrawl) via an encrypted local USDC wallet (Sep 26).
- claude-mood — Webcam/mic mood-detection plugin that feeds user-frustration signals to Claude as context (Sep 26).
- waitwise — Replaces Claude Code's spinner verbs with custom or Claude-generated content (Sep 26).
- theCity — Native macOS app visualizing live Claude Code sessions as a 3D city (Sep 27).
- jauvex — Electron app unifying Claude Code, Codex, and Grok sessions with two-way voice control (Sep 27).
- worktable-dev — MCP-based shared workspace where multiple agents/humans read and continue linked documents (Sep 25).
- Grabbit — Agent-first screenshot API with an MCP server for Claude/Cursor runtimes (Sep 26).
- Sitcom Flavour — Novelty plugin that flavors Claude Code chat replies with sitcom-style humor (Sep 25).