AI library

Weekly summary

Week 2026-W42

18 October 2026

The week in AI

  • Gemini 4 Argon goes public and GA (2026-10-13): Google's most capable model is on the paid Gemini API at $2 / $10 per 1M tokens, doubling after 2026-11-30. It has a 1M shared input-output limit and implicit caching only. Its safety report says it solved all 12 simulated cyberattacks, so Google applied its strongest safeguards. Penalising chain-of-thought made shortcuts silent rather than removing them. (launch, safety, model page)
  • METR embedded at Anthropic: six METR staff get employee-like internal access and can publish without approval, under a new three-tier scheme for evaluator access. (source)
  • Capabilities can be provably hidden: a paper hides capabilities in model weights behind a cryptographic trigger that every tested elicitation method missed, which limits what capability evals can prove. (source)
  • Claude Code moves to stateless MCP: from 2.1.292 local servers use the 2026-07-28 protocol by default, and from 2.1.299 remote connectors do too. 2.1.299 also compacts idle sessions before the prompt cache expires, and hooks can now fail closed. (2.1.292–296, 2.1.299–300)
  • Anthropic replaced git: with over 30,000 agent-driven merges a day, it built Scion, a server-side source control system. (source)
  • More models without open weights, and some with: Meta's Muse Wonder agent models are API-only. Mistral's Timestral time-series models ship open weights under a Modified MIT licence. (Muse Wonder, Timestral)

Library maintenance

  • Links: 75 checked, 0 dead, 0 moved, 0 blocked; no action needed.
  • Consolidated 14 topic pages updated this week:
    • models/claude: Current state cut from 9 to 4 entries. Glasswing, Cyber Mission, Genesis, Singapore, and evaluator facts moved into Key concepts, and Haiku 5.5 benchmarks and 4.x retirement dates into How to use it.
    • models/gemini: Current state cut from 9 to 4 entries. The overloaded "Current lineup" bullet was split into a lineup list and a separate Gemini 4 Argon entry, and Gemini 2.5 access limits, the Nano Banana deprecation, and price notes moved into the body.
    • practices/security: Current state cut from 7 to 4 entries. New "Defensive programmes" and "Frontier cyber capability" concepts hold the Cyber Mission and Argon results. The hidden-capabilities paper moved into Supply chain.
    • tooling/coding-agents: Current state cut from 7 to 4 entries. Scion figures, GovTech adoption, and Claude Code surfaces moved into Key concepts, and the duplicate Haiku-default entry was merged.
    • models/gpt: Current state cut from 7 to 4 entries. The GPT-6.1 Sol cache price, image-bug advice, Ultrafast residency, and GPT-Live moved into the body.
    • models/open-weight: Timestral, EmbeddingGemma 2, and Muse Wonder details moved into the body. The unconfirmed Qwen 3.7 Max report moved to Open questions.
    • practices/cost-and-caching: Current state cut from 6 to 3 entries. The Argon caching note and the Sonnet cache-cut date conflict moved into the price table.
    • frameworks/landscape: six release entries folded into the framework descriptions, with the version table kept as the current state. The answered Vercel question was removed and a DSPy 3.5 question added.
    • tooling/mcp, practices/evals, and fundamentals/reasoning-models: spec deprecations, the hidden-capabilities paper, Thinking Inertia, and per-model thinking controls moved into the body. The superseded 2026-10-04 reasoning entry was removed.
    • agents/memory, agents/agent-patterns, and fundamentals/training-and-fine-tuning: checked; already short (formatting fix only).
  • No stale pages: the oldest last_reviewed date is 2026-10-04, 14 days ago.
  • Inbox empty. No duplicate tags found. No open questions were newly answered beyond those the daily runs had already closed.