OpenAI opened global preview access to GPT-5.6 (codename Sol), plus Terra and Luna, ahead of a public launch Thursday, and Sam Altman kept it characteristically short: "happy building." The pre-launch reactions were unusually concrete for a tease.

Theo said it fixes what was wrong with GPT-5.5, Riley Brown got weeks of early access and called Codex "3x more fun", Justin Sun made it a daily driver for subagent delegation, and Codex CLI already added Sol/Terra/Luna as selectable models.

The most useful coding-agent conversation moved past "which model is smartest" to how to wire several together. Peter Steipete's one-liner — make Codex the workhorse under Fable 5 — drew 3,000+ likes, and the pattern repeated everywhere: orchestrate with Fable, route the grunt implementation to cheaper subagents.

running your Codex sub inside Claude Code to cut ~60% of tokens, framing it as "Fable orchestrates, Sol implements", and the still-unsolved Claude-to-Codex context handoff. Meanwhile SemiAnalysis showed ~95% of their Claude Code tokens are cache hits.

The open field widened on several fronts at once. Google shipped the Gemma 4 technical report for its E2B–31B multimodal models and, separately, on-device inference: Gemma 4 now runs fully offline in React Native apps via Vulkan and MLX delegates.

Gemma 4 went fully on-device via React Native, MiniMax is reportedly prepping a 2.7-trillion-parameter open model, and Alibaba's new open coding model reportedly set an 82% SWE-Bench Verified record under Apache 2.0.

Several items converged on one uncomfortable theme: the agent is now the vulnerability. The Hacker News flagged an attack that turns AI coding assistants into botnet installers by registering a repo name a model is likely to hallucinate, then hiding prompt-injection instructions inside it and waiting for agents to fetch the wrong resource.

"GitLost" tricks GitHub's AI agent into leaking private repos, alongside heavier classic bugs — a 15-year-old Linux kernel 0-day used in a full chain and a CVSS-10 ColdFusion path-traversal PoC — plus a new open-source red-teaming CLI for the defenders.

The local-AI thread shifted from "can you run it" to "does it pay for itself." The viral framing: people mocked buying three Mac Studios to run models at home — until a Fable outage and GLM 5.2's arrival sent those same machines reselling for 4× their price.

antirez pushed back on the "you can't run it locally" limits with dual M5 Max rigs, a $599 Mac mini got pitched as a private server to kill an API bill, and one used GPU reportedly cut a coding bill 93%.

A day after Muse Image climbed the image board, Meta's new video model entered the Text-to-Video Arena at #3 with a 1459 score — ahead of Alibaba's HappyHorse 1.0 by 30 points and ranking above Grok Imagine, putting Meta squarely into the open-ish video-model race.

The feed's runaway favorite wasn't a model — it was a website that lets you snip clippings out of old magazines from the Internet Archive (or your own PDFs) and assemble them into collage art. Simple, tactile, and squarely in the "beautifully made side project" lane, it pulled 14k+ likes.

For the tinkerers: "clawcodex" is a from-scratch, MIT-licensed port of the Claude Code agent loop in pure Python (230k+ lines) that its author claims runs roughly 230× cheaper. Whether the multiplier holds up in practice, an open, hackable reimplementation of the loop is a useful thing to be able to read and modify.