Claude Report — 2026-10-08

🚀 Anthropic Official

Claude Haiku 5.5 (Oct 7)

Launches as the new default Haiku model (claude-haiku-5-5) with a 1M token context window, 128K max output, and adaptive thinking with an effort parameter — Anthropic's cheapest and fastest model yet, costing roughly 75% less to run than Haiku 4.5. Scores jump sharply on agentic benchmarks (OSWorld 2.1: 72.4% vs 15.7% for Haiku 4.5; Terminal-Bench 4.0: 39.2% vs 0.0%). Pricing is $0.10–$0.50/Mtok input and $0.50–$2.50/Mtok output (rates roughly 5x higher once a prompt exceeds 100K tokens); manual budget_tokens thinking now returns a 400 error, a breaking change from Haiku 4.5. Ships across the Claude API, Bedrock, Google Cloud, and Microsoft Foundry.

Sonnet 5.5 prompt-cache price cut (Oct 7)

Cuts Claude Sonnet 5.5 prompt-cache read price from $0.20 to $0.10 per million tokens (0.05x base input instead of 0.1x), cutting the cost of long-context agentic workflows that lean on caching.

API credits for Max and Team plans (Oct 7)

Adds a monthly shared dollar pool usable on the Claude API, Managed Agents, the Agent SDK, and the Console playground (not interactive Claude Code): $100 for Max 5x, $200 for Max 20x, and up to $500 pooled for Team ($20/seat Standard, $100/seat Premium), non-rolling.

Browser and Computer Use SDK toolsets (beta) (Oct 7)

Adds Browser Use and Computer Use tool base classes to the Python and TypeScript SDKs — subclass one, implement one method per action, and the SDK runs the tool loop, URL/file policy checks, and approval callbacks against your own automation backend.

Managed Agents: web tool domain restrictions (Oct 7)

Applies allowed_hosts to web_search and web_fetch: non-matching web_fetch URLs now return url_not_allowed, web_search silently omits restricted-host results, and session creation fails with a 400 if an enabled tool's allowed_domains isn't covered by allowed_hosts.

Managed Agents: web_fetch URL security hardening (Oct 7)

Restricts web_fetch to URLs that already appeared in the session's user messages, prior web_search results, or prior web_fetch results — URLs seen only in Claude's own output, the system prompt, attachments, or tool outputs are rejected with url_not_in_prior_context, closing an indirect-injection exfiltration path.

Compliance API — unified Claude chats (Oct 8)

Extends Compliance API chat endpoints to also return chats from the unified Claude experience, in beta for Claude Enterprise using existing Compliance Access Keys.

Models API — server-tools capabilities (Oct 6)

Adds capabilities.server_tools to GET /v1/models, reporting whether each model supports web search and code execution tools, plus a top-level capabilities.code_execution flag for code calling other tools mid-execution.

Models API — thinking-disabled capability (Oct 5)

Adds capabilities.thinking.types.disabled so callers can check before sending thinking: {type: "disabled"} to a model.

Claude Code v2.1.294 (Oct 8)

Fixes prompt/agent hooks written as plain-language instructions (e.g. "Block commands that...") sometimes allowing what they should block, and improves how Stop/SubagentStop instruction-style hooks are judged so Claude stops early less often.

Claude Code v2.1.293 (Oct 7)

Makes Claude Haiku 5.5 the default Haiku model in the CLI and adds agentType to the subagent status-line payload plus isDeferred on $.tool.register for mods. Fixes a wide batch of issues: Claude re-doing work it had already finished after context compaction, an HTTP MCP memory leak, lost queued messages when backgrounding a session, /model effort-cycling wraparound, Claude in Chrome disconnects under /tui, login/session-folder read glitches in the footer and agents view, and several Slack ("Claude Tag") and Code Review polish fixes.

Claude Code v2.1.292 (Oct 6, carried from yesterday's window edge, confirmed in range)

Adds a --marketplace <source> flag to one-step-install a plugin from a new marketplace, and an effort parameter on the Agent tool to tune sub-agent effort level; ships a prompt.autocomplete hook, prompt caching for $.model.complete, and agent.spawn visibility for mods.

claude-agent-sdk-typescript v0.3.292 → v0.3.294 (Oct 6–8)

Adds agent_id/parent_task_id/run_id fields for tracking subagent and resumed-task lineage, typed sections/notes on ListAgents output, read_at/arrived_at on ReadNotifications, and a subagent_type field on background_tasks_changed entries; fixes subagents in auto permission mode wrongly using the classifier instead of canUseTool. v0.3.294 reaches parity with Claude Code v2.1.294.

claude-agent-sdk-python v0.2.164 (Oct 6, no new release since)

Bundles Claude CLI 2.1.292 and hardens CI with an egress-firewall runner, network allow list, and an enforcement-check workflow; raises the PyPI per-file size guard to 250 MiB.

anthropics/sandbox-runtime (Oct 5–7)

Ships srt-proxy, a standalone HTTP proxy with an external decider callback; adds per-command network allow lists with deny reasons surfaced through the ask callback; terminates TLS in-process on the tunnel socket instead of per-tunnel listeners; and cuts release v0.0.79.

anthropics/claude-quickstarts (Oct 6–7)

Adds a computer/browser toolsets quickstart exercising the new SDK Browser/Computer Use classes, and updates the daily-brief agent quickstart to run on Claude Sonnet 5.5.

Claude for Google Workspace — public beta (Oct 6, no further evolution since)

Opens a Claude sidebar in Google Docs, Sheets, and Slides on all paid plans, with "ask before edits" or "accept all edits" modes.

Claude Cowork — cloud sandboxes by default (Oct 6, no further evolution since)

Moves new Cowork tasks on Pro/Max to per-session cloud sandboxes by default, removing the local-execution toggle.

Cyber Verification Program expansion (Oct 6, no further evolution since)

Splits into three access tiers (Defense, Red Team, Specialized) with tier-appropriate reduced safeguards on Opus 5.5, Sonnet 5.5, and Mythos 5.1.

Claude Startups program expansion (Oct 6, no further evolution since)

Adds a free year of Claude Team, $1,000 in API credits, and up to $45,000 in partner offers for eligible startups via a new "Claude Startup Stack."

🔌 Claude Code Plugins

Install-count cache is unavailable this run, so top-10-by-installs ranking cannot be confirmed. Falling back to repo-folder activity in anthropics/claude-plugins-official for the window:

security-guidance — installs unconfirmed

Ships 2.0.10 → 2.0.11 (Oct 7): stops killed hooks from leaving stale git index copies and filling disk, after the 2.0.9 → 2.0.10 disk-cleanup fix already reported yesterday.

figma — installs unconfirmed

Bumps to v2.2.127 (Oct 7) via an upstream content sync.

salesforce-development — installs unconfirmed

Bumps to v2.3.0 (Oct 7) via an upstream content sync.

Unchanged in window: aws-core/aws-agents/aws-data-analytics (last bump Oct 2, outside window).

🛠️ Skills

None of the top 5 updated in window.

Unchanged in window: anthropics/skills (last commit Oct 5, already reported yesterday), kharmanskyi/open-steps (last release v0.4.8, Oct 5, already reported), addyosmani/agent-skills (no commits), K-Dense-AI/scientific-agent-skills (only an automated security-scan-report commit, Oct 5, not a feature change), cloudflare/security-audit-skill (no commits).

🤖 Agents & Subagents

davila7/claude-code-templates (Oct 7)

Fixes the jev-skill-typeahead mod's AbovePrompt bands to stack instead of hiding each other, and improves the product-strategist agent with a model field, sourcing discipline, and agent-integration hooks — new changes beyond yesterday's jev-skill-typeahead/payment-gateway-integrator work.

nicobailon/pi-subagents (Oct 6–8)

Adds provider display in subagent model listings, a one-line-per-run async TUI layout, and an OSC 7501-based subagent run-state reporting channel; bundles a tmux inspector plugin and fixes several subagent/workflow-resume race conditions — substantial activity beyond the v0.76.1 release already reported yesterday.

Unchanged in window: wshobson/agents (no commits), VoltAgent/awesome-claude-code-subagents (last commit Oct 5, already reported), anthropics/commerce-agents (dormant since Aug 31).

🔗 MCPs & Integrations

upstash/context7 (Oct 7–8)

Requires authentication on the hosted MCP endpoint, retries MCP Registry publishing until npm serves the version, and adds an Antigravity plugin for Context7 — new since yesterday's MCP server-card report.

oraios/serena (Oct 6–8)

Adds an EmmyLua language server option and prefetches language-server dependencies in the CLI, continuing from yesterday's TypeScript/Angular/PHP/Kotlin classification fixes.

DeusData/codebase-memory-mcp (Oct 7)

Merges a cross-repo status feature and fixes workspace-manifest entry resolution and Laravel route-handler edge cases, beyond yesterday's team-artifact-bundle work.

ChromeDevTools/chrome-devtools-mcp (Oct 7)

Adds an env var to enable the third-party tool category and fixes non-deterministic aggregate IDs in heap snapshots, beyond yesterday's name-truncation fix.

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

💡 Community — Workflows & Ideas

Piebald-AI/claude-code-system-prompts (Oct 6, no new release since)

Logged v2.1.292's internal system prompt growth; no new tagged release through v2.1.293/294 as of this run, so no further evolution to report.

thedotmack/claude-mem (Oct 6)

Ships three patch releases (v13.34.0–13.34.2) for its Claude Code memory-persistence tool, fixing a worker cache-invalidation regression, DSH (daemon shared-host) record handling on failed uninstalls, and restoring a Cursor trial flow — a concrete community-built long-term-memory layer for Claude Code sessions.

📰 Quick Mentions

• Claude Desktop v2.26454.2 — bundles Claude Code 2.1.293 with no additional user-facing changes (Oct 7).
• Hacker News: Claude Haiku 5.5 discussion — mixed reception; praise for speed/cost but concern over adaptive-thinking defaults and a contested 100K-token pricing cliff.
• The Hacker News — Cyber Verification Program coverage — recaps the three-tier access expansion alongside Project Glasswing's 129,000+ verified vulnerability count.