Fable 5 Subscription Caps Slash, Open Weights Close Cyber Gap, China Launches WAICO

📰 News

• Claude Fable 5 stays on Max/Team—but limits get cut hard — From July 20: bonus usage ends, regular caps drop 33%, and Fable runs at 50% of that reduced pool. Pro and Team Standard effectively lose plan access (one-time $100 credit, then API rates). Competition from GPT-5.6 Sol and Chinese pricing killed the “API-only” exit, but capacity is still the bottleneck. → The Decoder

• Open-weight cyber skills lag closed frontier by ~4–7 months — UK AISI: GLM-5.2 and DeepSeek V4-Pro hit levels closed models held months ago, at a fraction of the cost, with safety rails that fall off easily. The defender “prep window” is shrinking. → The Decoder

• China launches World AI Cooperation Organization (WAICO) — Shanghai HQ, 29 founding states, no Western sign-ons. Xi pledges 5,000 Global South training slots and pushes back on export-control framing—clearest bid yet for a parallel AI order. → The Decoder

• Navy AI playbook: slow adoption beats imperfect alignment as the bigger risk — “AI-first fleet,” Bits2Effects cycle, Mean Time to Effect. LLMs and agents on ships even when comms die. Sits in a world where GenAI.mil hit ~1.5M daily users. → The Decoder

• GPT-5.6 Full Access wipes home directories — No sandbox, $HOME overwrite path. OpenAI: it shouldn’t, but it did. Agent default permissions remain the scary part. → The Decoder

• Meta’s spare AI compute may land Anthropic first — Talks for up to ~$10B over two years (NYT). Claude Code demand needs GPUs; Zuckerberg already floated selling excess capacity. → The Decoder

• Apple’s lawsuit lands on OpenAI’s IPO clock — Partnership fight is now listing and valuation risk, not just product drama. → TechCrunch · podcast

• Hugging Face discloses AI-agent-driven intrusion — Malicious dataset hit loader + template injection; autonomous agent swarm moved laterally. Forensics on commercial APIs hit safety blocks—so they finished on local GLM-5.2. → HF

🔥 GitHub Trending

• anthropics/cwc-workshops — Code with Claude workshop kits: model picking, multi-agent skills/MCP, Managed Agents, memory, eval loops.

• github/copilot-sdk — Multi-language SDK (Python, TS, Go, .NET, Java, Rust) to embed Copilot CLI’s agent runtime in your apps.

• Nutlope/hallmark — Anti-AI-slop design skill: audit / redesign / study.

• PrismML-Eng/Bonsai-demo — Local 1-bit and ternary Bonsai demo—27B vision, tools, 256k; phone-scale footprint claims.

• HenryNdubuaku/maths-cs-ai-compendium — Intuition-first open textbook for maths/CS/AI, plus an MCP server for coding agents.

▶️ YouTube

• Claude Just Revealed AI’s Biggest Problem — Two Minute Papers | Quick take on Anthropic coding-assistance and skills research.

• Did Kimi K3 really beat Fable? — Matthew Berman | Beyond the headline benches—did it actually win?

• Kimi K3 just beat FABLE — Berman short | K3 highlight reel.

💬 Community

• Claude keeps Fable 5 on subscriptions — Willison: Sol/K3 made “API-only Fable” untenable. Still not on the $20 Pro plan. Fablepocalypse postponed.

• LLM cliché highlighter — Toy (Fable-vibe-coded) that flags ten LLM writing tells—“no fluff, no jargon,” etc.

• Quoting Kimi K3 — After refusing to leak its system prompt: “Is there something I can actually help you with today?”

• Verifiable AI inference — Proving a response really came from the claimed model.

• Inventing ELIZA — How the first chatbot shaped expectations (open-access PDF).

• Full-pipeline inference for MiMo-V2.5 — Xiaomi’s stack-level inference optimization notes.