An Insane Week for AI Agents — GLM 5.2, Codex Skills, Claude, and Cursor

The biggest story of the week was Z.ai's GLM 5.2. It matches GPT 5.5 and Opus 4.8 in capability while costing roughly 1/5 to 1/6 of frontier models. The whole thing is open-source, and integrating it into Cursor takes seconds.

The author maps the current frontier as Fable on top, GPT 5.5 close behind, and Opus trailing slightly. The real shock is that GLM 5.2 has now pulled level with that group. The single-frontier-leader narrative is collapsing in real time — both benchmarks and live use show the open-source side genuinely catching up.

On the Codex side, 'record and replay' is a serious step forward. Codex's computer use feature lets you record a screen workflow and register it as a skill. After that, Codex can call the skill and execute the workflow autonomously.

The live demo recorded 'upload my latest video to Typefully as a draft' as a skill, then asked Codex to run it via computer use. Recordings up to 30 minutes are supported, so even long workflows can be captured. Typefully's 512MB upload limit blocked the actual file in the demo, but the click-through, navigation, and upload flow on Comet executed correctly. The distance between 'record once' and 'automate from now on' just got dramatically shorter.

Looking forward, more Chinese open-source models are coming, and Gemini is expected to announce both a new model and a direction for its super app. Strategic alignment across the big labs is the thing to watch next. The author argues we've already been in the 'super app era' for 100 days.

The sharpest comment goes to Google. As a content creator, the author can't recommend Antigravity vs AI Studio vs Jules vs Gemini Desktop because there's no clear answer. Google has too many products, none of which feels like the super app. Picking *one* and committing is the prerequisite for Google getting back into real competition — that's the takeaway.