Claude Report — 2026-10-10

🚀 Anthropic Official

Claude Haiku 5.5 (Oct 7)

Launches as Anthropic's fastest, cheapest small model: 1M-token context, up to 128K output tokens, first Haiku with adjustable effort, multimodal (text+image), priced at $0.10/$0.50 per million input/output tokens (~75% cheaper than Haiku 4.5). Available on Claude Platform, AWS, Google Cloud, and Microsoft Foundry, plus Claude.ai for Free/Pro/Max/Team/Enterprise.

Claude Developer Platform — Oct 7 release notes

Cuts Sonnet 5.5 prompt-cache read price from $0.20 to $0.10 per million tokens (0.05x base input price); adds monthly API credits to Max and Team plans; adds Python/TypeScript SDK classes for browser-use and computer-use toolsets; and hardens Managed Agents web tools — allowed_hosts now gates web_search/web_fetch, and web_fetch can only retrieve URLs that already appeared in the session (blocks fetching URLs injected via tool output or system prompt).

Compliance API enhancements (Oct 8)

Lets Enterprise orgs retrieve unified Claude Experience (Chat+Cowork) sessions via the existing Compliance Access Key (beta), and adds Word-file export for Claude Docs through the Compliance API (beta).

Dynamic Workflows for Claude Managed Agents (beta) (Oct 9)

Lets a lead agent write and run a "workflow" — a program coordinating many agents in phases, executed server-side in the background — via the managed-agents-2026-04-01 beta header and workflows: {type: "enabled"} config. Supports up to 1,000 agents and 64 concurrent threads per run, tracked through workflow_run.* session events; internal benchmarks show workflows catching 66/70 planted bugs across 116K lines vs. 14–27 for a single agent.

Claude Code v2.1.296 (Oct 9)

Adds CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL to pin every workflow-spawned agent to one model, autoCompactWindow for subagent frontmatter/--agents definitions, allow_large on the Read tool, and CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS; fixes managed-settings hooks that denied calls with "continue": false without ending the turn, headless sessions with plugin MCP servers, Claude apps gateway sign-in regression from v2.1.295, and Windows PowerShell/stdio-MCP issues; raises the MCP tool-description limit from 2,048 to 4,096 characters.

anthropic-sdk-typescript sdk-v0.133.0 / sdk-v0.132.1 (Oct 8–9)

Adds types for Chat/Cowork unified analytics metrics and for Managed Agents workflows, multiagent configuration, and thread-status filtering; v0.132.1 fixes sending requests with an empty path parameter.

anthropic-sdk-python v1.13.0 / v1.12.1 (Oct 8–9)

Mirrors the TypeScript SDK: adds Chat/Cowork analytics types and Managed Agents workflows/multiagent/thread-status support; v1.12.1 adds CI changelog checks and pins the uv.lock package version.

claude-agent-sdk-typescript v0.3.296 (Oct 9)

Reaches parity with Claude Code v2.1.296: adds claude_code_version to the initialize control response and autoCompactWindow to AgentDefinition, fixes sandbox option merging with inline settings, and raises the MCP tool-description limit to 4,096 characters.

anthropics/skills — claude-api dynamic-workflow quickstarts (Oct 9)

Syncs the official claude-api skill with the Dynamic Workflows release and adds four new quickstart templates (bug-hunter, performance-tuner, explainer-video-maker, watchlist-scanner), bringing the total to 13 and teaching managed-agents-onboard to scaffold each one's environment.yaml/session.yaml.

🔌 Claude Code Plugins

Plugin tracking skipped this run — marketplace install-count cache unavailable, per instructions. No commit-level sweep of the top 10 official plugins was performed today.

🛠️ Skills

anthropics/skills

Adds four dynamic-workflow quickstart templates to the claude-api skill (see Official section above for detail).

addyosmani/agent-skills (Oct 7–10)

Merges a dozen+ PRs tightening the skill set: a frontmatter linter that now ignores comments and nested metadata keys, fixes preserving reference-heading anchors and source details during prose review, and documentation clarifying .cursor/ behavior and lifecycle slash-command limits in Copilot; still unreleased past tagged v0.6.12 (Oct 3).

Unchanged in window: kharmanskyi/open-steps (v0.4.8, last bumped Oct 5), K-Dense-AI/scientific-agent-skills (last release Oct 1), cloudflare/security-audit-skill (no activity since Sep 14).

🤖 Agents & Subagents

davila7/claude-code-templates (Oct 9–10)

Ships a security fix for a Console Bridge injection vulnerability (GHSA-4jm9-m3fr-9jpx, v1.29.7), adds a CI build for the Anthropic OSS Scanner Dockerfile with a threat-model writeup, and adds HeyReagent MCP to the marketplace listings.

nicobailon/pi-subagents v0.77.0 (Oct 8–10)

Cuts a new release after two days of fixes: shares idle-wake state with pi-intercom to stop concurrent wake failures, terminates detached foreground children during runtime replacement, records parent-session linkage on subagent runs, and scopes interrupts to the requesting run only.

Unchanged in window: wshobson/agents (last commit Oct 4), VoltAgent/awesome-claude-code-subagents (last commit Oct 5), anthropics/commerce-agents (last commit Aug 31).

🔗 MCPs & Integrations

upstash/context7 (Oct 9)

Ships ctx7@0.6.0 with a --language option on ctx7 docs to prioritize code examples in a chosen programming language, and ctx7@0.5.16 hardening MCP OAuth JWT validation (strict signature/audience checks) and flagging broken docs via isError.

oraios/serena (Oct 8–9)

Follows up on the ~50% file-traversal speedup merged this week with regression tests, a unified project-file-gathering refactor, .git directory ignore handling, and an Erlang language-server race fix.

DeusData/codebase-memory-mcp (Oct 9–10)

Merges a TF-IDF vocabulary-comparison fix for semantic search (vocabulary vs. token-position mismatch) and a new document-linking feature connecting index status to source documentation.

ChromeDevTools/chrome-devtools-mcp (Oct 9)

Adds support for debugging WebMCP tools directly in third-party developer tools (#2900); the --allowUnrestrictedPaths=false CLI fix flagged yesterday is still unreleased (latest tag remains v1.10.1, Sep 23).

Unchanged in window: graygnatconsole/mcp-audit-tool (no commits since Sep 26).

💡 Community — Workflows & Ideas

Dynamic Workflows Move to the Server (Oct 9)

Argues the real shift in Anthropic's new feature is architectural, not just scale: execution moves from client to server, running independent of user interaction, which raises new problems around cost control, observability, and ensuring tool calls stay idempotent across up to 1,000 parallel agents.

Claude's dynamic workflows find 66 of 70 planted bugs (Oct 9)

A synthetic benchmark across 116,000 lines of code shows the multi-agent workflow consistently finding 66/70 planted bugs over three runs, versus 14–27 for a single agent — concrete evidence for when orchestration beats a single long-running agent.

Piebald-AI/claude-code-system-prompts v2.1.296 (Oct 9)

Logs a smaller but notable follow-up to Oct 8's record jump: +560 tokens adding managed-agent quickstart templates (bug hunter, explainer video maker, watchlist scanner) and expanded self-hosted-runner troubleshooting guidance to the Claude Code system prompt.

franzenzenhofer/big-arrow-on-the-screen (Oct 9)

A Claude Code/Codex skill shipping a single Swift CLI binary that lets agents draw a big, click-through arrow and label on the Mac screen to point at what a human should click — no daemon, telemetry, or AI inside; hit 300+ points on Hacker News.

📰 Quick Mentions

• HN: Claude Haiku 5.5 — top-engagement thread (1,044 points) on pricing vs. GPT-6 Luna and real-world token costs (Oct 8).
• Claude Dynamic Workflows Beta: Can It Beat One Agent? — independent hands-on test of the new Managed Agents workflows beta (Oct 9).