For me, it’s not a question of whether I like it, nor where we might move in the future, but purely what is proven currently. In my experience, as in my evals, any model, including notably Fable 5 and GPT-5.6 Sol can and indeed do produce code that does not consider many edge cases, can be hard to maintain (especially Sol though it’s output has become markedly better with recent changes to Codex) and sometimes introduce bugs with visibly apparent issues. Existing test coverage doesn’t necessarily prevent such issues when changes or additions are made. The idea that models tend to gravitate towards the house style without major CI based restrictions also has not born fruit in my book. More, I have found even the frontier models amazing at utterly deviating from established standards and tests, what lint, tc and co don’t catch the model will take to introduce code that feels and reads like it does not fits with the rest.

Models also still can have major task adherence issues, so even a perfectly well scoped, well considered prompt that covers all situations and coverage can, in many scenarios lead to output users may not expect or notice without reading code. Seen it a lot, just lost some blood and tears to a very straight front end implementation task that I ended up applying manually cause both Sol and Fable struggle to copy a few dozen HTML/JS/CSS snippets properly into a slightly obscure codebase, despite the reference being fully featured and even using the same token scheme, etc…

We also see such issues in what labs deliver in their own software products. We are not at the stage where a human shouldn’t sign off on code that they have tested and understood. A hypothetical future, I don’t really care for, we’d notice when labs stop shipping easily apparent bugs, though even then, I’d still feel there is a while in between that and not needing to assess the output anymore.

There is, in my mind, no need for this “narrative shifting”, overtone window (don’t get me started on the idea of applying that to what should be a purely technical discussion), convince the unbelievers nonesense that some spend their time on. If models could provide code that didn’t necessitate review, there’d be no one needing to argue that case. If and when they can, we’ll know it, but not because of those espousing this currently.