Claude Report — 2026-09-02

🚀 Anthropic Official

Claude Fable 5.1 and Claude Mythos 5.1 (Sep 1)

• Fable 5.1 ships GA across the Claude API, AWS, Google Cloud, and Microsoft Foundry; same $10/$50 per MTok pricing as Fable 5, but cache-read cost drops 75% to $0.25/MTok (up to 45% total cost savings on agentic workloads).
• 1M-token context window, 128K max output, always-on adaptive thinking.
• Mythos 5.1 is the same base model with reduced safeguards, restricted to vetted "Project Glasswing" cybersecurity/life-science partners (reportedly including Microsoft, Apple, AWS, Cisco, Nvidia) — not publicly released.
• Cited results: ~50% hit rate on protein-binder design, high-resolution Venus elevation mapping, up to 2.5x GPU kernel speedups.
• Both models require ≥30-day data retention (not available under zero data retention) and carry Anthropic's text watermark plus C2PA content credentials.

Enterprise Frontier Safeguards (Sep 1)

Introduces "Enterprise Frontier Safeguards," combining zero-data-retention with automated (not Anthropic-reviewed) misuse monitoring whose logs stay in customer-controlled AWS/Azure/GCP infrastructure; opt-in, no added cost, rolling out fall 2026 across Claude Code, Enterprise, Platform, Bedrock, Google Agent Platform, and Microsoft Foundry. Co-developed with 100+ customers including Goldman Sachs, Morgan Stanley, Mastercard, Visa, Salesforce, Stripe, and KPMG.

Claude Code v2.1.257 and v2.1.258 (Sep 1)

• v2.1.257 (17:53 UTC): adds Fable 5.1 as the default Fable model (1M context), adds Time Format/timeZone settings, adds a new "Containment Escape" auto-mode rule that stops auto-approving cloud metadata-credential fetches, egress evasion, and cross-tenant reach, and adds env-var-based subagent model control.
• v2.1.258 (22:33 UTC): fixes Claude Code failing to launch on macOS 12 (Monterey) — a regression from v2.1.255 — and fixes remote/scheduled sessions erroring with "user messages must have non-empty content" after a re-sent permission approval.

Claude API changelog (Sep 1)

• Adds claude-fable-5-1 and claude-mythos-5-1 model IDs with the specs above.
• tool_choice values any/tool now return a 400 error on the new models (auto/none still work).
• Tightens thinking-block validation: blocks preserved only by the producing model or newer, 400s on altered history for accounts created after Aug 31.
• Adds three betas: per-message "effort" control via a system-role message, turn-scoped system messages via clear_at: "next_user_message", and a new thinking.display: "updates" value.

Agent SDK TypeScript v0.3.257/v0.3.258, Python v0.2.150/v0.2.151, Anthropic SDK Python v1.3.0 (Sep 1)

• Agent SDK TS adds thinkingTokens to ModelUsage, fixes resourceLinks on MCP tool-result messages, fixes browser-SDK streaming on engines lacking native Symbol.dispose, and adds a detail option to Query.getContextUsage(); both point releases (v0.3.257, v0.3.258) track the same-day Claude Code CLI bumps to 2.1.257 and 2.1.258.
• Agent SDK Python mirrors the same CLI-version bumps (v0.2.150 → CLI 2.1.257, v0.2.151 → CLI 2.1.258).
• Root Anthropic SDKs (Python v1.3.0, TypeScript sdk v0.123.0) add beta user-profile fields, org compliance-settings and memory-store/toolset schema updates, and fix an AWS base_url/aws_region bug under skip_auth plus a credential-file leak into non-Node bundles (TS).

🔌 Claude Code Plugins

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

🛠️ Skills

tt-a1i/archify

Ships two license/attribution fixes — preserves license provenance in packaged distributions (Sep 1, #265) and bundles third-party notices (Sep 2, #267); stars climb to ~42,983 (+2.3k since Aug 31).

agentskillexchange/skills

Adds a new "Pronounce Developer Jargon" skill to its catalog (Sep 2) on top of its continuing near-hourly automated sync.

anthropics/skills

Lands a major update to the claude-api skill (Sep 1) documenting Claude Fable 5.1/Mythos 5.1, GA status for the Files and Skills APIs, a new cost-optimize subcommand, and expanded Managed Agents memory-store/domain-allowlist docs.

Unchanged in window: K-Dense-AI/scientific-agent-skills (only a cosmetic README trim and an automated security-scan-report commit, still v2.65.0), alirezarezvani/claude-skills (main branch unchanged since Aug 26; a merge landed only on the unmerged dev branch Aug 30).

🤖 Agents & Subagents

nicobailon/pi-subagents

Cuts v0.63.0 (Sep 1) and lands a same-day-plus-one watchdog feature (Sep 2, unreleased) adding consequential findings, child cadence, a diff tool, and launch rules.

anthropics/claude-agent-sdk-typescript and claude-agent-sdk-python

See Anthropic Official above — each ships two coordinated point releases (v0.3.257/v0.3.258 and v0.2.150/v0.2.151) on Sep 1, tracking Claude Code's same-day CLI bumps.

Unchanged in window: VoltAgent/awesome-claude-code-subagents (README-only sponsors update Sep 1, already reported), AMAP-ML/LongHorizon-Harness (last push Aug 20).

🔗 MCPs & Integrations

modelcontextprotocol/servers

Merges CI/test hardening (Sep 2) — unifies the git_log output schema and drops raise_exceptions from server.run, and lets fork PRs get labeled via pull_request_target.

github/github-mcp-server

Reactivates with seven merged PRs on Sep 1, including expectedHeadSha pinning for merge_pull_request and a fix for silently dropped issue labels.

upstash/context7

Rebuilds its mcpb packaging as a self-contained bundle published on release (Sep 1, #3099).

DeusData/codebase-memory-mcp

Continues its high-cadence merge streak into Sep 1 — Go cross-package field dispatch, Perl grammar v1.2.1, Windows CI guard fixes, and an index-worker memory budget.

Unchanged in window: modelcontextprotocol/typescript-sdk (last change Aug 31, DPoP merge already reported), punkpeye/awesome-mcp-servers (Sep 1 sponsorship credit already reported yesterday), modelcontextprotocol/python-sdk (last merge Aug 28), oraios/serena (last fix Aug 30).

💡 Community — Workflows & Ideas

Breaking Claude Code Opus 5 Auto Mode (Aug 31)

Top HN story (397 points) demonstrates a bypass of Claude Code's Opus 5 auto-mode safety gating; plausibly the trigger behind the new "Containment Escape" auto-mode rule Anthropic shipped the next day in Claude Code v2.1.257.

Security Flaw Enables Untrusted Repos to Execute Code in Claude Code & Others (Sep 2)

Discloses a git-hijack vector letting an untrusted repository trigger code execution when opened by Claude Code and other coding agents.

awesome-claude-code new resources (Sep 2)

Merges four new community tools in one day: TermaGITchi (a git-state-driven status-line "creature"), tmux-claude-status-tabs (event-driven tmux session indicator), faf-cli (auto-generates CLAUDE.md/AGENTS.md from a single source file), and a Dynamic Workflow Design Patterns guide/plugin.

SlideOps (Aug 31)

New Agent Skill generates HTML slide decks from repo code and adds a dependency-free freshness checker that flags slides whose cited code has since changed, repairing only the affected slides.

Pairmark (Sep 2)

New benchmark harness races Claude Code against Codex head-to-head on the same repository to compare coding-agent performance.

📰 Quick Mentions

• Claude Code's +50% weekly usage promo expires (Aug 31) — the temporary boost lapsed after three extensions; the Sept 14 "permanent" limit increase is calculated off the pre-boost base, netting roughly a 17% cut versus today's boosted level.
• anthropics/claude-plugins-official bumps its Security Plugin to v0.11.0 (Sep 1) — part of routine marketplace-repo maintenance alongside a frontend-design skill copy fix.
• Ration (Aug 30) — local-only macOS menu-bar app shows a usage ring auto-calibrated from your own Claude Code session history.
• MaCcyP (Aug 31) — fork of the Maccy clipboard manager adds an "Agents view" so Claude Code can push snippets without overwriting the system clipboard.
• Claude Code Arcade (Aug 31) — Node.js daemon renders animated arcade games directly in the Claude Code status line via Unicode glyph sprites.
• HN discussion: Claude Code now appends a session link to every commit (Aug 31) — thread on the behavior change embedding Claude session links in commit messages.
• Skills vs. MCP tools for AI agents: When to use which (Sep 2) — comparison of Claude Agent Skills versus MCP tool-calling design tradeoffs.