Anthropic keeps Model 2 in-house, Grok serves word salad, China closes the gap

📰 News

Anthropic’s strongest model, “Model 2,” stays internal — About 1.5 AECI points above Mythos 5. Used for coding, data, research agents. Claude already writes most of Anthropic’s production code. No public release planned. → The Decoder
Grok keeps answering in gibberish — Grok Lite on grok.com only. Ask for a PDF, get “match it without and your they…” for paragraphs. A fresh chat usually clears it. The X account is fine. → TechCrunch
China has caught the frontier — Kimi K3, GLM-5.3, and Qwen3.8-Max sit near the top of broad benches. What’s left: reliability and cyber. A model lead isn’t a moat anymore. → The Decoder
35% of pages since ChatGPT look AI-written — Pew + Pangram. .com is ~10× .edu/.gov. Em dashes and “it’s not X, it’s Y” are up too. → TechCrunch
Sutton: synthetic data is a “big mistake” — The world is vastly bigger than any sim. Agents should learn from their own experience. LLMs are maybe 20–25% of intelligence. → The Decoder
Ramp launches Router, an OpenRouter rival — OpenAI, Anthropic, DeepSeek, xAI, and more. Routing free through 2026. Logs inputs/outputs for a year unless you opt out. → TechCrunch
Adobe Firefly adds Gemini Omni Flash — Royalty-free music, voiceover, SFX. Video, audio, and images in one prompt. → The Decoder

jundot/omlx — Menu-bar LLM server for Mac. Hot RAM + cold SSD KV cache, built for local Claude Code.
mukul975/Anthropic-Cybersecurity-Skills — 817 agent skills for security work. Not an Anthropic product. Mapped to ATT&CK and NIST.
mattpocock/skills — Small, composable skills for real engineering, not vibe coding. Plugin or copy the files.
obra/superpowers — Spec first, TDD plan, then subagents grind for hours. Claude, Codex, Grok, and more.

11 INSANE Use Cases for Grok Bot — Matthew Berman. Email, calendar, DoorDash, coding — then it tidies your computer.

smolvm as a sandbox for untrusted Python & JavaScript — Simon. Claude Fable 5 hit no-KVM, so it ran the tests on GitHub Actions instead.
Conceptual integrity and counting lines of code — Agents mint code fast. Software grows Winchester Mystery House rooms.
Quoting Jeremy Morrell — LLMs make extensions cheap; sandboxes make shipping them safe.