№ 1502026-09-23

Claude Report — 2026-09-23


🚀 Anthropic Official

Introducing Claude Opus 5.5

Launches the first model in the new 5.5 family, matching Claude Fable 5.1 on most tasks while cutting run costs 40% versus Opus 5. 1M-token context window, 128k max output, always-on adaptive thinking (cannot be disabled), 30% faster generation with an optional 2.5x "fast mode," and 85% fewer boundary-circumvention attempts on Anthropic's automated behavioral audit. Pricing drops 20%: $4/$20 per Mtok (input/output, was $5/$25), cache reads $0.20 (was $0.50), cache writes $5 (was $6.25); fast mode runs $8/$40. Adds preserved-thinking anti-distillation safeguards and expanded biology/cybersecurity verification programs.

Claude API release notes — Opus 5.5 breaking changes & inline tools (Sep 22)

  • Thinking can't be disabled for Opus 5.5: thinking: {"type": "disabled"} now returns a 400 error; use the effort parameter instead.
  • tool_choice types any and tool return 400 for Opus 5.5; only auto with strict tool use is supported.
  • Computer use now requires the computer_toolset_20260801 toolset (the older computer_20251124 returns 400 on the Claude API and Google Cloud, though it still works on Bedrock).
  • Fast mode (research preview) becomes available for Opus 5.5 on the Claude API.
  • New beta: tools can be defined inside mid-conversation system messages via the inline-tools-2026-09-15 header, letting you add tools or update schemas without touching the top-level tools array or invalidating the prompt cache; a paired mcp-client-2026-09-15 header enables MCP toolset definitions with a new mcp_tool_listing response block.

Claude Code v2.1.280 (Sep 22)

Makes Claude Opus 5.5 the new default Opus model and ships a broad reliability/UX pass: better fullscreen mouse support, stronger MCP controls (including a CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH override for the 2,048-char cap), cleaner dialogs, VS Code session archiving, and a hook change where agent-type hooks no longer run for PermissionRequest events (command/http hooks now handle those, with an explicit error pointing there).

Official SDKs ship Opus 5.5 support + inline tool definitions (Sep 22–23)

  • anthropic-sdk-python v1.8.0 — adds claude-opus-5-5 support, inline tool definitions, MCP tool-list pinning (beta), a shared evaluated_permission enum, and a Python 3.13 open-stream crash fix.
  • anthropic-sdk-typescript v0.128.0 — same feature set, shipped alongside same-day companion releases for the Vertex, Bedrock, Google Cloud, Foundry, and AWS SDKs.
  • claude-agent-sdk-python v0.2.158 — adds ClaudeAgentOptions.verbatim_prompts, which delivers user messages to the CLI exactly as written, blocking @path file expansion and slash-command dispatch so untrusted text can't trigger unintended file reads or command execution; bundles CLI 2.1.280.
  • claude-agent-sdk-typescript v0.3.280 — adds fireReason for scheduled task-notification origin, a verbatimPrompts option, an alpha readMcpResource() for MCP Apps UI resources, and gives askSideQuestion() visibility into current-turn context.

🔌 Claude Code Plugins

Plugin tracking skipped this run — official marketplace install-count cache unavailable.

🛠️ Skills

addyosmani/agent-skills

Extends validators to enforce the same frontmatter requirements on commands as on skills, rejects invalid YAML in command frontmatter, checks reference links inside skills' own reference directories, restores security-hardening rules that had been inadvertently dropped, and pilots a Claude-plugin evaluation for code-review-and-quality (Sep 20–23).

Unchanged in window: eugeniughelbur/obsidian-second-brain (last commit Sep 21, already reported), affaan-m/ECC (last commit Sep 21, already reported), agentskillexchange/skills (only automated sync commits plus the already-reported concept-to-story skill), cloudflare/security-audit-skill (last commit Sep 14), K-Dense-AI/scientific-agent-skills (only an automated scan-report commit Sep 21), tt-a1i/archify (only cosmetic README/branding/website commits), ayghri/i-have-adhd (no commits since Sep 19).

🤖 Agents & Subagents

nicobailon/pi-subagents

Ships v0.71.0 (Sep 23) restoring activated-tools/workflow-child evidence, exposing workflow terminal proof, stripping inherited Git routing environment variables, keeping readable child-session names, and discovering symlinked prompt workflow files — roughly 30 commits across the window following v0.70.1 on Sep 21.

Unchanged in window: VoltAgent/awesome-claude-code-subagents (only a README edit Sep 21), anthropics/commerce-agents (no commits since Aug 31).

🔗 MCPs & Integrations

ChromeDevTools/chrome-devtools-mcp

Releases v1.10.0, finishes the "migrate MCP SDK to v2" refactor, extracts a BrowserManager class, fixes JavaScript-dialog handling during input-tool actions, sets default pageSize/pageIdx for get_css_styles, and adds a CI smoke test for bundled binaries without node_modules (Sep 21–23).

oraios/serena

Bridges external language-server adapters through its registry (Sep 22) on top of last window's MCP SDK v2.2.0 upgrade, plus fixes for Dart rootUri/rootPath omission, aborting memory renames on read-only memories, and .csproj files excluded by project ignore settings (Sep 21).

upstash/context7

Ships a new Search API: TypeScript SDK integration, documentation, and quickstart config updates land across Sep 20–23.

DeusData/codebase-memory-mcp

Keeps shipping daily fixes: a CI file-classifier fallback for large diffs, a Go CGO pseudo-import fix, a Cypher non-numeric LIMIT fix, a .NET XML-extension fix, and new AI-assisted-contribution guidelines (Sep 20–22).

Unchanged in window: github/github-mcp-server (last commit Sep 16), modelcontextprotocol/ext-skills (no commits in window).

💡 Community — Workflows & Ideas

davila7/claude-code-templates

Adds a jev-skill-suggestion mechanism that injects chosen skills into context while hiding the selection step from the model, plus a new workspace-orchestration skill and a Mods trending/catalog filter for the CLI's component browser (Sep 20–22).

📰 Quick Mentions

← back to archiveend of edition № 150