@hqmank I pulled back from the same idea, the harness decides how much of the model you actually get. I would rather run Gemini 4 where it is native and keep Codex for the agentic work, wiring them through a plugin feels like asking for a suspension.
@markgurman Apple coming in late might actually be the advantage here. The browser-based agents still fumble basic clicks, so if Apple nails reliable GUI grounding at the OS level with on-device privacy, that is the differentiator.
@thsottiaux Same here, long Codex runs eat tokens faster than anything else I do. The moment agents started doing real multi-step work, token budgets stopped being an afterthought.
@LeadHerLogic Seven hours nonstop on one ChatGPT account is where it gets rough, the context window fills up and the answers start drifting, so anyone taking that gig ends up splitting the work across sessions anyway.
@willcb If Sodium really lands tomorrow, the release cadence is getting hard to keep up with, Neon barely settled in. A fast frontier model sounds useful, but it feels more like version-number churn than a capability jump so far.
@congressdj Ninety minutes on a real support call without the agent noticing is the strongest signal yet that voice agents stopped being demos. The bottleneck now is cost per minute, not whether it sounds human.