A week after the GPT-5.6 launch, the conversation moved from "which tier is best" to the reality of living inside Codex all day. OpenAI's Thibault Sottiaux reset usage limits across Codex and the new ChatGPT Work — twice in one day — a tell that people are hitting caps hard. The Codex team's Reddit AMA put a number on it, and the most useful tweets were practical: how to rein in Sol's token appetite, and how not to get burned.

5M+ people now use Codex weekly, a config recipe to tame Sol's token burn, a full agentic setup dumped in the open, a blunt "don't run agents in yolo mode", and the Codex CLI still nagging about GPT-5.5.

The open-weight story stopped being about catching up and became about parity and permanence. Maziyar Panahi showed GLM-5.2 running an autonomous structural-biology loop — driving the Mol* viewer, having Qwen3-VL critique its own renders, and refining until the inhibitor sat in the pocket — the kind of multi-tool science people assumed needed a frontier closed model. The benchmarks backed the vibe, and under it all ran a note of urgency.

GLM-5.2 trading blows with Sonnet 5, the "is 5-10x more worth it?" question, a call to hoard weights while you can, China's open-weight strategy taking the perf-per-cost lead, and the GLM founder's AGI claim.

Claude Code kept compounding. Roundtable showed it going multiplayer — humans and agents working in the same terminal session — while a quiet point release shipped two dozen CLI changes and the skills ecosystem got real. The maturity cut both ways: as tooling made it easier to build with, researchers were rewarded for breaking it.

2.1.207's 24 CLI changes, Anthropic's four agent "loops", a tool that auto-installs the right skills, and a $3.7k bounty for a .git sandbox escape.

Security Twitter had a loud, actionable week. Socket caught the popular jscrambler npm package shipping a malicious preinstall hook — a live supply-chain compromise of a package developers pull without thinking. The AI-agent attack surface widened too, and the classic stack didn't get a break either. The throughline: the tools and pipelines developers trust are now the attack.

Ghostcommit hiding prompt-injection in images, a WhatsApp message turning OpenClaw into a RAT, a CVSS 9.8 Windows HTTP.sys RCE, and a critical Zimbra stored-XSS fix.

The recurring "AI video is almost there" refrain flipped to "I couldn't tell." The strongest evidence was mundane on purpose: a 4K Seedance 2.0 product render with believable skin pores and fabric, and short cinematic clips that read like real camera work. The other half of the story was control — turning one-off generations into repeatable, directed pipelines.

a 15-second cinematic made entirely in Seedance, a clean character swap with Seedance + GPT Image 2, a Seedream 5.0 vs GPT Image 2 comparison, and directed paper-cut animation from Midjourney styles.

With open weights at parity, the conversation moved to iron. Builders showed off private inference rigs — a Dell 7740 packed with four RTX Pro 6000s, an eight-GPU Arc Pro B60 box — framed as work you stop renting from someone else's machine, while others argued a single 4090 already replaces most of a monthly Claude bill. There was a reality check, and a counterpoint from the software side.

an 8-GPU Arc Pro B60 inference box, a 4090 replacing most Claude spend, an honest homelab reality check, and a sprint making Gemma 4 run 5x faster on one A10G.

Scott Ullrich, who co-founded pfSense, revealed he's spent the last year building its successor — an unusually direct pitch to the self-hosting and homelab crowd that's grown frustrated with pfSense's closed development. It was one of the most-shared self-hosted tweets of the day.

Charlie Holtz turned a real engineering pain into an agent task: Conductor was Mac-only, which made testing manual and local, so he asked Claude Fable to port it to Linux and open the pull request itself. A clean, concrete example of an agent doing end-to-end cross-platform work rather than a demo.