1312026-09-04

Claude Report — 2026-09-04


🚀 Anthropic Official

Claude Fable 5.1 and Claude Mythos 5.1 (Sep 1)

  • Ships as Anthropic's most advanced coding/knowledge-work models, with early scientific-research capability; model IDs claude-fable-5-1 and claude-mythos-5-1.
  • Sets a 1M-token default context window, 128k max output tokens, and always-on adaptive thinking.
  • Scores 52.6% on agentic scientific research (vs. 24.7% for Fable 5), 55.8% on Terminal-Bench, 60.9% on multidisciplinary reasoning.
  • Cuts cache-read pricing 75% to $0.25/MTok (0.025x base input, vs. 0.1x on other models); standard rate stays $10/$50 per MTok input/output.
  • Restricts tool_choice to auto/none only — any and tool now return 400 errors; use strict tool use or structured outputs instead to guarantee schema-conformant inputs.
  • Preserves thinking blocks only for the producing model or newer ones; new accounts (created Aug 31, 2026+) get a 400 error when replaying blocks after a system-prompt/tools/message change.
  • Adds three betas: per-message effort switching mid-conversation, turn-scoped system messages (clear_at: "next_user_message"), and streamed thinking progress updates.
  • Requires 30-day data retention for both models (no zero-data-retention option outside Enterprise Frontier Safeguards); text carries Anthropic's watermark, code-execution media gets C2PA Content Credentials.
  • Restricts Mythos 5.1 to vetted US organizations via the Cyber Verification Program and a government-partnered Life Sciences Verification Program; cybersecurity safeguards cut false positives 60%, biology safeguards fire 85% less often on benign requests.
  • Launches across Claude.ai, the Claude API, AWS Bedrock, Google Cloud, and Microsoft Foundry/Azure.

Enterprise Frontier Safeguards (Sep 1)

Introduces a system, co-developed with 100+ customers, that stores enterprise data in infrastructure the customer controls instead of Anthropic's systems while keeping automated safety monitoring active; rolls out in phases starting this fall, with eligible customers able to use Fable 5.1 under zero data retention in the meantime.

Claude Code v2.1.260 (Sep 3–4)

  • Adds a diff panel that opens beside the conversation in fullscreen mode, showing uncommitted changes live as Claude edits, toggled with /diff.
  • Adds prompt-cache-miss diagnostics that surface likely causes (tool-definition or system-prompt changes).
  • Adds /reload-plugins support for headless sessions and a text-based /advisor command for desktop and headless environments.
  • Fixes file-permission rules containing parentheses being dropped as invalid, and zsh commands with hidden command substitutions being auto-approved incorrectly.
  • Fixes Bedrock model discovery failing when corporate root certificates exist only in OS stores, and managed-settings loading for Enterprise/Team users on legacy API keys.
  • Fixes prompt caching on Fable 5.1 not covering context attached after tool results, and SDK-provided MCP servers missing from initial turns.
  • Fixes /rewind not restoring properly from backup files, stale file-read tracking after rewinds, and agent-team transcript loss during extended API retries.

Agent SDK TypeScript v0.3.260 (Sep 3)

Adds user_message_uuid on thinking_tokens system messages and four remote-session latency-breakdown fields (first_content_frame_ms, first_stream_post_ms, first_stream_post_ack_ms, first_stream_post_wall_ms); fixes managedSettings.disableAutoMode: "disable" being silently dropped and rewindFiles() falsely reporting success when checkpoint backups are missing; tracks Claude Code CLI 2.1.260 parity. The Python SDK remains at v0.2.152 (tracking CLI 2.1.259), with no release yet matching 2.1.260.

🔌 Claude Code Plugins

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

🛠️ Skills

anthropics/skills — frontend-design update

Updates the frontend-design skill to steer away from generic design defaults (Sep 3, PR #1713).

agentskillexchange/skills — BulkPublish skill

Adds a new "BulkPublish" social-publishing skill to the auto-synced catalog (Sep 3, 17:08 UTC), on top of its routine hourly regeneration syncs.

Unchanged in window: tt-a1i/archify (last commits Sep 1–2, already reported), K-Dense-AI/scientific-agent-skills (last push Sep 2, already reported), alirezarezvani/claude-skills (no change since Aug 26).

🤖 Agents & Subagents

nicobailon/pi-subagents v0.65.0

Ships fail-closed behavior for child-agent model resolution and cleans up native-subagent edge cases (Sep 4), building on the prior watchdog release.

anthropics/claude-agent-sdk-typescript

See Anthropic Official above — ships v0.3.260 (Sep 3) with latency-breakdown fields and two bug fixes; the Python SDK counterpart has not moved past v0.2.152.

Unchanged in window: claude-agent-sdk-python (still v0.2.152), VoltAgent/awesome-claude-code-subagents (README-only churn, no content change), AMAP-ML/LongHorizon-Harness (no push since Aug 20).

🔗 MCPs & Integrations

github/github-mcp-server

Adds an Agent Plugins 1.0 package (#3169) and fixes sanitizer content-boundary and Markdown-fidelity bugs on read surfaces (Sep 3).

upstash/context7

Adds authentication and usage tracking for the Claude Code plugin integration and ships a version-packages release (Sep 4), following Sep 2's new VS Code, Devin, and GitHub Copilot CLI setup targets.

DeusData/codebase-memory-mcp

Merges fixes for Windows guard-daemon isolation, strict environment-integer parsing, scoped return * matching, and dotted TypeScript import resolution, plus CI hardening for macOS/Node smoke tests (Sep 4).

oraios/serena

Adds FreeBSD platform detection and removes unneeded language-server platform restrictions, persists removed registered projects, reaps orphaned language-server child processes, isolates Hardhat state for Solidity on macOS, and fixes losing memory content on interrupted writes (Sep 3).

modelcontextprotocol/typescript-sdk

Fixes OAuth resource indicators being altered instead of preserved exactly, and surfaces underlying network errors via Error.cause on client probe failures (Sep 3).

Unchanged in window: modelcontextprotocol/servers & python-sdk (no new commits since the memory-hardening batch and OAuth-issuer validation already reported), punkpeye/awesome-mcp-servers (last commit Sep 1).

💡 Community — Workflows & Ideas

No relevant news today.

📰 Quick Mentions

← back to archiveend of edition № 131