Weekly summary

Week 2026-W46

15 November 2026

The week in AI

  • Claude Fable 5.2 and Mythos 5.2 arrived at $8 / $40, 20% below 5.1, with account-bound thinking and a new prefix check to watch for in API code. Fable 5.2 is now Claude Code's default Fable model. (source)
  • Gemini 4 Carbon (GA) and Gemini 4 Hydrogen fill out Google's Gemini 4 line: Carbon is the top model at $2 / $10 until 2026-11-30, then $4 / $20, and Hydrogen is the cheap, fast model at $1.50 / $7.50. (Carbon, Hydrogen)
  • Update mcp-server-git to 2026.11.11 or later. Older versions ran commands from a repository's own .git/config, even on read-only calls (High). The filesystem and fetch servers got fixes the same day. (source)
  • Claude Code 2.1.319 to 2.1.323 added MCP tasks in the background, permissions.denyInsteadOfPrompting for unattended runs, plugin version ranges, and an experimental Code Mode. They also fixed many permission gaps, including wildcard Bash allow rules that skipped prompts. (2.1.319, 2.1.323)
  • Anthropic's "The components of agent memory" found that strong models do better when told what memory is for than when given rules about what to save. (source)
  • MCP moves to a six-month spec release cycle; the next version is 2027-03-30. (source)

Rising on GitHub

  • anthropics/skills (194,734 stars, +5,785 this week): Anthropic's official skills, including Word, PDF, PowerPoint, and Excel document skills. /plugin marketplace add anthropics/skills, then /plugin install document-skills@anthropic-agent-skills.
  • google/elephant-goldfish-model (19,307 stars, +5,141): drafts a design doc with full context, then has fresh, context-free subagents review it before you build.
  • coreyhaines31/marketingskills (60,399 stars, +1,291, MIT): skills for copywriting, SEO, conversion, analytics, and email, built on a shared product-marketing skill. npx skills add coreyhaines31/marketingskills -a claude-code.
  • D4Vinci/Scrapling (103,058 stars, +3,620): a scraper exposed as an MCP server, with saved logins and screenshots, that strips hidden prompt-injection text by default.
  • mksglu/context-mode (34,112 stars, +456): keeps large tool output in a local sandbox and passes the agent only what it needs, so long sessions use less context.

Skipped: dozens of identical "GitHub-Copilot-app" and "blender-ai-copilot" repos with around 400–670 stars in one to three days (likely spam), and xerj, whose benchmarks are self-reported and whose quickstart turns off authentication.

Library maintenance

  • Links: 205 checked, none dead. Updated one moved URL (Claude Help Center article on AI-generated content labels). 4 blocked (OpenAI, Physical Intelligence): left as they are.
  • Consolidated: coding agents (10 → 3 Current state entries, folded into the body) and Gemini (9 → 3, with upcoming dates in one entry). Also trimmed Claude, MCP, skills, frameworks, cost and caching, security, and open-weight models by removing stale GitHub trend notes and entries already in the body. Fixed a broken Sources list in cost and caching, and brought the model lineup on LLMs up to date (Fable 5.2, Gemini 4 Carbon and Hydrogen).
  • Stale pages reviewed: tokens and context (limits table updated; checked against Anthropic's models overview, which says the tokenizer introduced with Opus 4.7 fits about 555K words in 1M tokens, down from about 750K), transformers (MoE sizes, GLM-5.3 sparse attention), and training and fine-tuning (post-training-only releases, MXFP4 post-training). No new source notes were needed.
  • Housekeeping: the inbox is empty, and no tags overlap.