The coding fight moved from model quality to the client layer. OpenAI's Tibo announced inference optimizations passed through as price cuts for GPT-5.6 on Codex and ChatGPT Work — "no nerfing" — and separately flagged a Codex-OSS fork crossing 60k stars. Meanwhile xAI's Grok Build started letting you resume Claude Code, Codex, and Cursor sessions from a single prompt, so the agents increasingly become interchangeable front-ends over whichever model is cheapest. Underneath it all, users keep asking to escape the terminal, and pressing Anthropic that reliability — not just intelligence — is what earns loyalty.

a Codex-OSS fork hit 60k stars, Grok Build resumes Claude/Codex/Cursor sessions, the push to get out of the terminal, and an open letter on trust vs. intelligence.

This week's open-weight story was compression and cheap silicon, not new leaderboards. Brian Bell showed a lossless, read-in-place compression that shaved 423GB off GLM-5.2 (1,403 to 980GB, 753B weights, bit-exact, no quantization) so the full model fits in less VRAM. Google put Gemma 4 31B on Cerebras at 1,500+ tokens/sec, and another team teased a faster GLM-5.2 quantization. On the hardware side, people are repurposing $60 AMD BC-250 boards with 16GB of PS5-grade GDDR6 to cancel monthly AI subscriptions.

Gemma 4 at 1,500+ tok/s on Cerebras, a new GLM-5.2 quantization, a $60 local-AI GPU hack, and the case that consumer GPUs catch up by 2029.

A strong day for actionable security. Nebula Security won kernelCTF again with a fresh 0-day (another $100k bounty). Naval Postgraduate School researchers published a genetic-algorithm attack that hides prompt injections in binary string variables to make LLM-powered reverse-engineering agents — Ghidra plus GhidraMCP — misread executables, a new attack surface as decompilation goes agentic. On the classic side, a Samba %J print-command injection yields unauthenticated RCE with a public PoC, and a popular Chrome extension was caught exfiltrating browsing history.

binaries weaponized against RE agents, a Samba unauth-RCE CVE with PoC, a 1.6M-install extension exfiltrating history, and a 2023 injection baked into a model's weights.

The video-model leaderboard flipped: Google's Gemini Omni Flash debuted at #1 on Artificial Analysis's Text-to-Video and Image-to-Video boards, edging out ByteDance's Seedance 2.0. ByteDance answered from the product side — Seedance 2.5 went live on Jimeng with first 720p looks and a beta showing an "ultra-long video" mode (up to 180s) and 4K/30-second generation — while Dreamina's Seedance 2.0 Mini rolled out to the US and globally. The frontier is now long-form and 4K, not just clip quality.

Seedance 2.5 went live on Jimeng, it added a 180s / 4K beta, and Dreamina's Mini went global.

Designers are drawing a line between models that draft and models that ship. Psalter's side-by-side argued that GPT-5.6 Ultra generates a workable-but-generic site while Fable 5 makes the same content look studio-shipped — same underlying task, very different polish. Practitioners echoed it: warnings against burning usage treating Claude Design as a brainstorming toy, a full landing page rebuilt from a single hero screenshot, and Fable 5 paired with Higgsfield to collapse the cost of an animated marketing site. The question is shifting from "can AI design?" to "which model for which stage?"

stop treating Claude Design as a toy, a landing page from one screenshot, and Fable 5 + Higgsfield vs. the agency.

Agent "skills" went from hobby to first-party. itshan surfaced Apple's official Agent Skills exported from Xcode 27 — SwiftUI, UIKit modernization, Swift Testing, C bounds-safety, and security hardening, packaged for AI coding agents. Around it, builders shared their own: Obsidian's CEO open-sourced his Claude Code skills for anyone to clone, and Ruben Hassid published a practical anatomy of why a skill does or doesn't fire. The ecosystem is standardizing on skills as the unit of reusable agent capability.

Obsidian's CEO open-sourced his skills, and a breakdown of why skills fire.

Raycast released an official extension for mymind, letting you search, save, upload, and organize your mymind library through its API without leaving the launcher — a tidy pairing for anyone who lives in both.

Prime Intellect released verifiers v1, an overhaul of its environment stack for agentic RL and evals that decomposes environments into a taskset, a harness, and a rubric — infrastructure for training and grading agents as harness design becomes the critical layer.

An update to the open pi-computer-use agent brings it to parity with Codex's recently upgraded computer-use, adding niceties like a ghost cursor to show intended actions — a sign the open computer-use tooling is keeping pace with the labs.