Claude Report — 2026-10-01
- Anthropic launches <a href="https://www.anthropic.com/news">Claude Sonnet 5.5</a> (Sep 28) — 30% faster and cheaper than Sonnet 5, live on API/Bedrock/Vertex/Foundry.
- Claude Code ships <a href="https://github.com/anthropics/claude-code/releases/tag/v2.1.284">v2.1.284</a> and <a href="https://github.com/anthropics/claude-code/releases/tag/v2.1.286">v2.1.286</a>; Sonnet 4.5 gets a Nov 30 retirement date.
- Base SDKs add Admin API enterprise controls and an <a href="https://github.com/anthropics/anthropic-sdk-python/releases/tag/v1.10.0">MCP tunnels beta</a> (v1.10.0/v1.11.0, Sep 30).
- <a href="https://github.com/anthropics/claude-plugins-official/commit/aa5654b7">claude-plugins-official</a> adds a new math-proof plugin and patches context7/imessage (Sep 30).
- Community ships <a href="https://github.com/mksglu/context-mode">context-mode</a> (context-window optimizer) and <a href="https://github.com/mvschwarz/openrig">OpenRig</a> (multi-agent harness), both trending Oct 1.
🚀 Anthropic Official
Claude Sonnet 5.5 (Sep 28)
Launches on the Claude API, Bedrock, Vertex, and Microsoft Foundry, claiming ~30% faster responses at up to 30% lower cost than Sonnet 5 for most workloads. The migration guide flags five breaking changes: a renamed thinking parameter, forced tool_choice now returning 400s, thinking blocks now tied to model/account, rejection of the computer_20251124 tool on API/GCP, and the advisor tool rejecting older models as advisors.
Claude Sonnet 4.5 deprecation (Sep 30)
Announces API retirement of claude-sonnet-4-5-20250929 for Nov 30, 2026, pushing teams toward Sonnet 5.5 migration.
Claude Code v2.1.284 (Sep 28)
Makes Sonnet 5.5 the default Sonnet model, adds USD spend-limit tracking, effort-slider keybindings, and a new /rate-limit-options command.
Claude Code v2.1.286 (Sep 30)
Adds permission-prompt counting ("2 of 5"), mouse support for list navigation, and fixes auth/credential handling, MCP connector/auth bugs, subagent transcript/worktree issues, VS Code extension bookmarks, Slack "Claude tag" channel spend limits, and a Windows trust-record case-sensitivity bug.
Agent SDKs catch up to CLI 2.1.286 (Sep 30)
claude-agent-sdk-python v0.2.163 bumps the bundled CLI and pins the CI model for issue triage; claude-agent-sdk-typescript v0.3.286 adds SDK MCP manifest support and fixes foreground-subagent tool access.
anthropic-sdk-python/typescript v1.9.0–v1.11.0 (Sep 28–30)
Adds the between_tools thinking type and GA cache diagnostics (v1.9.0, Sep 28); ships Admin API features for enterprise analytics, spend limits, RBAC, Plugins/Marketplaces management, and an MCP tunnels beta (v1.10.0, Sep 30); adds a list-spend-limits endpoint and Sonnet 4.5 deprecation support (v1.11.0, Sep 30). Matching releases ship for the TypeScript SDK plus Vertex/Bedrock/Google Cloud/Foundry variants.
claude-plugins-official — new math-proof plugin (Sep 30)
Adds a new math-proof plugin with "solo" and "siege" (multi-agent) skills for hard mathematics problems; the same day, context7 gets an OAuth-fallback fix and imessage (v0.2.0) is updated to surface every inbound image instead of just the first.
sandbox-runtime — Linux hardening (Sep 30–Oct 1)
Improves Linux glob-walk performance and CI hardening (Sep 30), then merges a fix for mandatory names resolution below a path (#629, Oct 1).
🔌 Claude Code Plugins
Install-count cache unavailable again this run — formal top-10-by-installs ranking is skipped per instructions. Unranked plugin-folder activity found in claude-plugins-official this window is covered above under Anthropic Official (math-proof, context7, imessage); no other plugin folder among the ~50 in the repo had commits in the window.
Unchanged in window: all other plugin folders in claude-plugins-official (e.g. agent-sdk-dev, code-review, feature-dev, pr-review-toolkit, security-guidance, skill-creator, asana, discord, github, gitlab, linear, playwright, terraform, and ~30 more) — none had commits in the Sep 28–Oct 1 window.
🛠️ Skills
anthropics/skills
Updates the claude-api skill to default to Claude Opus 5.5, adds Claude Sonnet 5.5, and adds new build-eval/hill-climb guide flows plus a preserved-thinking migration subcommand (Sep 29).
K-Dense-AI/scientific-agent-skills
Ships v2.70.0 (Sep 29), consolidating repo guidance from CLAUDE.md into AGENTS.md and adding tests for recognizing "unsupported size" disclaimers in generate-image.
Unchanged in window: kharmanskyi/open-steps (its v0.4.6 release was already covered yesterday; no further concrete evolution since), addyosmani/agent-skills (last commit Sep 26), cloudflare/security-audit-skill (last commit Sep 14).
🤖 Agents & Subagents
nicobailon/pi-subagents
Ships v0.74.0 (Sep 30): fixes /subagent-cost dropping workflow receipts, corrects resumed foreground-workflow round counting, and treats MCP clients' string "true" as boolean workflow: true.
wshobson/agents
Follows up its dangling-reference cleanup with a hol-guard pin repoint to upstream main and fixes for antigravity install paths/dead skill links (Sep 28-29).
Unchanged in window: davila7/claude-code-templates (only routine automated data syncs, no feature work since Sep 17), VoltAgent/awesome-claude-code-subagents (last commit Sep 21), anthropics/commerce-agents (last commit Aug 31, increasingly stale).
🔗 MCPs & Integrations
ChromeDevTools/chrome-devtools-mcp
Fixes a submodule-init build issue, patches a correctness bug rejecting named groups in URL patterns, and refactors worker logic into a new McpWorker class (Oct 1).
upstash/context7
Fixes peer-dependency version bumping to only trigger when a range is exceeded and switches AI-SDK tool tests to run through OpenRouter (Sep 30-Oct 1).
DeusData/codebase-memory-mcp
Merges fixes for an "unknown daemon treated as foreign" activation bug, a harness timeout race condition, and test-runner directory isolation (Sep 30).
oraios/serena
Stops the olb IDE instance on project shutdown for the olb-jvm backend, drops the suffix naming scheme for language backends, and improves REPL tool-call logging (Sep 30).
Unchanged in window: graygnatconsole/mcp-audit-tool (last commit Sep 26).
💡 Community — Workflows & Ideas
context-mode
Sandboxes tool output in isolated subprocesses for up to 98% size reduction (e.g. 315KB→5.4KB), persists session state in SQLite+FTS5 hybrid search so sessions survive context compaction, and uses hooks to enforce routing across 17 agent platforms.
OpenRig v0.6.0
Ships a local daemon/CLI/TUI/MCP server built on tmux that defines multi-agent topologies via a YAML "RigSpec" (pods/edges/continuity policies), letting Claude Code and Codex agents send/broadcast to each other and snapshot whole team configurations.
Hamel Husain — critiquing Claude Code's build_eval/hill-climb
Reviews Anthropic's new auto-eval workflow and flags design flaws: it commits to evaluators before reviewing data, validation prompts lack context, and checks get bundled instead of separated.
agent-dispatch plugin
Ships v1.0.0 (Sep 28) for scheduled, headless issue-to-PR automation with run-state management, followed by a v1.1.0 fix the next day.
ccpocket / Ciao
Launches a WebSocket-bridge mobile client that surfaces Claude Code/Codex agent questions and approvals as phone push notifications, letting users steer a running session remotely (HN Show, Sep 30).
📰 Quick Mentions
- Barclays scales Claude — official customer-adoption story on expanding Claude use across business functions (Oct 1).
- issue-to-pr v9.5.0 — adds stateful run resumption and tighter Codex-collaboration integration for autonomous PR workflows (Sep 29).
- tg-essentials v2.0.0 — renames/expands the tg-alerts plugin into broader Telegram account management (Sep 29).
- maximhq/bifrost — fast enterprise AI gateway with MCP support, trending with 8,496 stars (Oct 1).
- Hao0321/claude-skill-social-post — topic:claude-skill repo trending today, a possible future top-5 skill candidate alongside Storybloq.