Claude Report — 2026-10-17
- Claude Code ships v2.1.301 (Oct 15) and v2.1.302 (Oct 16): fork subagents on by default, idle-compaction toggle, Claude Design sunset notice.
- Claude Code Review exits research preview, goes GA for Team/Enterprise with ~3x faster reviews (Oct 15).
- Anthropic SDKs add a `speed` param and spend-cap visibility; Agent SDKs ship in lockstep with each Claude Code CLI bump (Oct 15-16).
- Claude Mythos 5 autonomously designs quantum circuits beating published state-of-the-art on 7 of 8 benchmarks, up to 17x fewer gates (Oct 16).
- Anthropic grants METR "employee-like" embedded access with veto-free publication rights for independent safety evaluation (Oct 15).
🚀 Anthropic Official
Claude Code v2.1.301 (Oct 15)
Adds a retry-watchdog env var (CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS) to end requests instead of waiting out distant usage-limit resets, a vim % bracket-match motion, multi-line mod text inputs, richer OpenTelemetry tool_decision fields (classifier_verdict, permission_mode, classifier_category), and VS Code LaTeX math rendering. Fixes fable/best model-alias errors by retrying the previous version, worktree deletion without warning on quit, Grep/Glob silently reporting no matches on search-tool failures, and stray terminal control characters on VTE/iTerm2.
Claude Code v2.1.302 (Oct 16)
Adds CLAUDE_CODE_AUTO_MODE_MODEL to pick the client-side auto-mode classifier model, an idle-compaction toggle in /config, and a deprecation notice that standalone Claude Design (claude.ai/design) closes December 14, 2026. Enables fork subagents (subagent_type: "fork") by default, defaults claude -p/Agent SDK sessions without a permission mode to default on third-party providers, cuts plugin/agent load time via fewer filesystem calls (especially on Windows), and fixes the Read tool loading entire images into memory (now refuses files over 128MB), prompt-cache misses on tools reloaded mid-session, and a dozen Claude-in-Slack bugs (HEIC/SVG/TIFF images not reaching Claude, Enterprise Grid DM routing).
Standalone Claude Design sunsets December 14, 2026 (Oct 16)
Anthropic is folding the standalone claude.ai/design site into Claude proper — it has worked inside any Claude conversation since September 16 — because running two parallel versions slows fixes for both. After the close date the URL redirects into Claude; chats and comments left in the standalone version are deleted and public project links stop working, though design systems can be migrated from the Artifacts page first.
claude-agent-sdk-python v0.2.168 / claude-agent-sdk-typescript v0.3.301 (Oct 15) / v0.3.302 (Oct 16)
Ship in parity with the Claude Code CLI bumps: the Python SDK bundles CLI 2.1.301, while the TypeScript SDK adds per-tool MCP permission policies (always_ask, always_deny) and BUILTIN_TOOL_NAMES exports in v0.3.301, then fixes structured-output error delivery and switches permissionMode defaults for third-party providers in v0.3.302.
anthropic-sdk-python v1.15.0 / anthropic-sdk-typescript sdk-v0.135.0 (Oct 15)
Adds a speed parameter and usage.speed field to the Messages API, adds between_tools to thinking-type model capabilities, and returns a Console organization's usage-tier spend cap as a read-only spend limit. The TypeScript SDK additionally fixes streams being left open when dropped unread.
Claude Platform API changelog (Oct 15)
Three entries: the Compliance API now exposes org members' "Instructions for Claude" and Cowork global instructions via a new GET /v1/compliance/organizations/{org_uuid}/user_preferences endpoint; Fast Mode exits research preview for Claude Opus 5.5 (no beta header needed, default rate limit on Start/Build/Scale tiers, priced at $8/$40 per MTok vs. $4/$20 standard); and Fast Mode for Opus 5.5 lands on Claude in Microsoft Foundry for Azure-hosted deployments.
Claude Code Review is generally available (Oct 15)
Exits research preview for Claude Team and Enterprise orgs on GitHub (including Enterprise Server/Cloud with data residency): reviews now run automatically on new PRs once an admin enables the GitHub App, finish about 3x faster, and offer Quick/Standard/Deep settings per repo. Eligible Enterprise orgs can claim $10,000+ in credits through January 15, 2027.
An update on independent evaluators (Oct 15)
Anthropic grants METR "employee-like access" — six staff on-site with company hardware and internal system access — under a three-tier evaluator model where METR, the top tier, can publish risk assessments without Anthropic veto (redacting only predefined categories like customer data). SecureBio covers bio/chem risk review, Accenture handles red-teaming, and talks with Hugging Face are in early stages; the arrangement mirrors embedded bank-examiner supervision and responds to the White House Accord on Super Intelligence.
Designing more efficient quantum circuits with Claude (Oct 16)
In a 53-hour experiment led by researcher Andi Gu through Anthropic's STEM Fellows Program, Claude Mythos 5 autonomously designed quantum circuits beating prior state-of-the-art on 7 of 8 computational-primitive benchmarks (chemistry, optimization, simulation, number theory, numerical computing) — the standout modular-inversion design used up to 17x fewer gates than earlier published approaches, with fewer qubits too.
🔌 Claude Code Plugins
Plugin tracking skipped this run — marketplace install-count cache unavailable, per instructions.
🛠️ Skills
None of the top 5 updated in window.
Unchanged in window: addyosmani/agent-skills (last commit Oct 10), anthropics/skills (last commit Oct 9), kharmanskyi/open-steps, K-Dense-AI/scientific-agent-skills (only an automated security-scan commit Oct 12, no feature change), cloudflare/security-audit-skill.
🤖 Agents & Subagents
davila7/claude-code-templates v1.23.0 (Oct 16)
Ships a new release on top of this week's mod/subagent-tuner additions to this 200K+-download CLI/component catalog that configures commands, hooks, MCPs, and agents for Claude Code.
Unchanged in window: VoltAgent/awesome-claude-code-subagents (last commits Oct 7), wshobson/agents (last push Oct 5), anthropics/commerce-agents, nicobailon/pi-subagents (v0.78.0 still latest, from Oct 13).
🔗 MCPs & Integrations
DeusData/codebase-memory-mcp v0.12.0 (Oct 15-16)
Ships the v0.12.0 release flagged as upcoming in yesterday's report, syncing packaging across npm, PyPI, Go wrapper, and Scoop manifests, with a CI fix giving the wrapper smoke test a Go toolchain and a deletable work tree.
upstash/context7 (Oct 16)
Repositions the project for AI-agent use throughout README, docs, and plugin messaging, removes deprecated Skill Hub CLI commands, and fixes a bug where rejected refresh tokens didn't prompt re-sign-in.
ChromeDevTools/chrome-devtools-mcp (Oct 16)
Lands 20 commits including a fix that discovers third-party developer tools before tool lookup, a smaller third-party CLI bundle for faster startup, and viewport-restoration fixes for Lighthouse runs that fail mid-emulation.
Unchanged in window: oraios/serena (last release v2.0.0 Oct 13, no newer commits Oct 16-17), modelcontextprotocol/servers (still v1.0.0 from Oct 11).
💡 Community — Workflows & Ideas
thedotmack/claude-mem v13.40.0 / v13.41.0 (Oct 16-17)
This persistent-memory layer for coding agents adds TeamBrain terminal integration (npx claude-mem teambrain join <invite-link>) and cloud work-state sync in v13.40.0, then layers on workflow skills for build/release/build-and-release operations and privacy-preserving usage telemetry in v13.41.0 — a fast-moving evolution of the memory-timeline-restoration fix reported earlier this week.
📰 Quick Mentions
- DE-CIX launches an open-source MCP server for natural-language network management (Oct 15) — built on the open IX-API standard, lets agents find connectivity options, book/scale services, and check status via any MCP-compatible client.