OpenAI's plan to hand everyone a coding agent leaves the hard part to the model

"Forgiving technical audience" is the load-bearing phrase in the whole conversation ^[7]. A forgiving audience is one that can tell a bad answer from a good one before it ships, and that already knows which parts of its job are worth handing to a machine. Neither property travels with the packaging. Sottiaux's account of the transition is about maturity of the model and breadth of distribution ^[7], not about replacing the judgment that Codex users brought with them from their day jobs.

That is where the design rule starts to bite. The stated approach is to get out of the way of the model, keep the product surface minimal, and let conversation carry the interaction rather than making people learn an application ^[8]. It is a coherent answer to the interface problem. It is not an answer to discovery, which TechCrunch billed as one of the interview's three subjects ^[12]. Buttons and menus are ugly, but they are also the inventory: they tell a first-time user what the thing does. Remove them and the only remaining teacher is the model itself, mid-task, to a user who does not yet know what to delegate. ChatGPT Voice growth is offered as evidence that natural interaction wins, without figures attached ^[9].

Ethan Mollick's framing, put to Sottiaux by TechCrunch, is the sharper version of the same problem: ChatGPT Work reaches for magic, while Anthropic's Claude Cowork puts A/B tests in front of you and makes you choose ^[11]. One product spends the user's attention, the other spends the model's confidence. Asked whether workers are ready for the second approach, the published answer begins "We definitely see that" and the excerpt stops ^[4]. So the strongest counter-argument on the table goes unrebutted in the text we have.

The cost-of-intelligence half is thinner still. There is one number: $20 a month, with Work inside the Plus plan ^[6]. The only capacity mechanism described anywhere in the piece is human and discretionary, since TechCrunch introduces Sottiaux as the person Codex users know for resetting token limits when the product hits a growth milestone ^[3]. No task budget or usage allowance for Work appears in the interview ^[13]. For an engineer on a company card, an opaque allowance is an irritation. For a $20 subscriber whose agent halts partway through a job they were not confident specifying in the first place, the allowance is the product experience, and they have no way to tell a limit from a failure.

Worth noting who owns both sides of this. Sottiaux's remit covers the metered API and the flat-rate consumer subscription, plus agent infrastructure and enterprise ^[1], reporting to Greg Brockman ^[2]. When TechCrunch raised the analyst view that OpenAI needs to own the application relationship with the user, the answer moved to utility and willingness to pay ^[10]. Fair enough as pricing philosophy. It leaves the question of who holds the relationship exactly where the question of who teaches the user was left: with the model.