AI News 2026-07-30

$ cat top-story.md

OpenAI Cuts GPT-5.6 Luna and Terra Prices by Up to 80%

Just three weeks after launching GPT-5.6, OpenAI is discounting its two cheaper tiers: Luna (fastest/cheapest) drops 80% to $0.20/$1.20 per million input/output tokens, and Terra (balanced, everyday-work tier) drops 20% to $2/$12. Flagship Sol keeps its price. The move lands as OpenAI faces fresh price/performance pressure from Claude Opus 5, Kimi K3 and other frontier competitors.

Why it matters: If any part of your pipeline routes through the OpenAI API, Luna and Terra just got meaningfully cheaper for high-volume tasks — worth revisiting your model-routing logic and cost comparisons against self-hosted Gemma.

// source: OpenAI · additional sources: CNBC, Axios

## 📦 Models & Releases

▸ Google Ships Gemini Robotics 2 With Whole-Body Humanoid Control

DeepMind's new robotics stack lets humanoids coordinate full-body movement (walk, crouch, manipulate) with finer multi-finger dexterity and lets different robot types collaborate on shared tasks, backed by a new agentic safety benchmark. Why it matters: A preview of where general agentic reasoning is heading beyond the chat window — the same embodied-planning techniques keep showing up in non-robotic agent harnesses too. // DeepMind Blog

## 🛠️ Tools & Repos

▸ openwork Is an Open-Source Alternative to Claude Cowork

A free desktop app (macOS/Windows/Linux) built on OpenCode that mirrors Cowork's "hand it a folder" workflow, but lets you bring your own model and provider. It's trending hard on GitHub, gaining close to 1,000 stars today. Why it matters: A drop-in open alternative if you like Cowork's UX but want to point it at your own self-hosted vLLM/Gemma endpoint instead. // GitHub Trending

▸ Agent-Reach Gives Coding Agents Free Web Search Across Reddit, X, YouTube and GitHub

A CLI that Claude Code, Codex and other agents can call to read and search Twitter, Reddit, YouTube, GitHub and more with zero API fees, automatically falling back between scraping methods when one breaks. Already near 63k stars. Why it matters: A no-API-key way to give your agents live web awareness instead of building your own scraping layer. // GitHub Trending

## 🤖 Agents & MCP

▸ book-to-skill Turns Any Technical Book Into a Claude Code Skill

Feed it a PDF/EPUB/DOCX and it extracts frameworks, patterns and per-chapter reference files as an on-demand-loadable skill, compatible with Claude Code, GitHub Copilot CLI and Amp — cutting token use 24-51x versus dumping the raw text into context. Why it matters: A concrete pattern for turning your own reference docs or papers into cheap, reusable Claude Code skills. // GitHub Trending

▸ last30days-skill Researches Any Topic Across Reddit, HN, X and Polymarket

A Claude Code skill that synthesizes a grounded trend summary from Reddit, X, YouTube, Hacker News and Polymarket for any topic — currently one of the fastest-growing repos on GitHub, past 55k stars. Why it matters: The same category of tool this newsletter's own virality-signal step leans on — handy for ad-hoc trend research inside a Claude Code session. // GitHub Trending

## 🏠 Self-Hosting & RAG

▸ vLLM Details Its Arm CPU Inference Optimizations

A deep dive into Arm CPU enablement work landing in vLLM, covering kernel and scheduling optimizations that meaningfully close the gap to GPU-class throughput for CPU-only serving. Why it matters: Relevant if any part of your Gemma/RAG stack runs on Arm hardware (Graviton, on-prem Arm servers) instead of GPUs. // vLLM Blog

## 🎬 Worth Watching

▸ Did Anthropic Just Kill the Indie Hacker?

Fireship's fast, opinionated take on how Claude Opus 5's coding proficiency squeezes the classic indie-hacker business model of thin SaaS wrappers around LLM calls. Why it matters: A quick gut-check on what frontier coding models mean for anyone shipping solo on top of AI APIs. // YouTube — Fireship

▸ Kimi K3 Just Broke the Economics of AI

A digestible breakdown of what Moonshot's 2.8T-parameter open-weight release means for inference costs and the open-vs-closed model economics. Why it matters: Good primer if you're weighing whether Kimi K3 is worth the ~700GB of GPU memory it takes to self-host. // YouTube — Two Minute Papers

▸ Open-Source-Palantir: World Monitor für alle

German-language walkthrough of World Monitor, the AGPL-licensed, self-hosted OSINT dashboard that aggregates 65+ data sources into one situational-awareness UI. Why it matters: A hands-on look at a fully self-hostable data-aggregation stack, if you enjoy running your own instead of paying for one. // YouTube — c't 3003

## 📰 Worth Reading

▸ AI Worming Through Word

A researcher demonstrates a self-replicating prompt-injection attack against Microsoft Word's Copilot: hidden instructions get copied into every document Copilot touches, turning each output into a new carrier. Why it matters: A concrete reminder to sanitize untrusted document content before any agent — including your own RAG ingestion pipeline — touches it. // simonwillison.net

▸ Discovering Cryptographic Weaknesses With Claude

Anthropic researchers let Claude Mythos grind for 60 hours (~$100k in compute) on cryptanalysis, turning up genuine mathematical flaws in HAWK and a weakened AES variant, and open-sourced a new CryptanalysisBench. Why it matters: A rare, well-documented case of a model doing real novel research rather than just summarizing it. // simonwillison.net

▸ The Economic Benefit of Refactoring

A hands-on study shows refactoring a 17k-line Rust file into smaller modules cut the input tokens an agent needs for the same change by 83% (159,564 → 27,360), because a well-organized codebase means the agent has to read less. Why it matters: Hard numbers to justify spending agent time on refactoring instead of only new features. // martinfowler.com, HN 120 pts

AI News — July 30, 2026