I published over 100 billion tokens to Codex from Cursor, using intense manual supervision for collaborative programming, without loops or goal-driven token burning.

The most impressive moment was the release of Claude Sonnet 3.7, which brought the fundamental programming experience to a whole new level.
Although subsequent models are becoming increasingly intelligent, it’s very clear that the marginal utility is diminishing.
And if we consider the ROI of tasks, this year's state-of-the-art models all fall short; the economic model of coding tokens is definitely doomed.
People don’t need models that are this smart,
but rather models like Cursor Composer, which are cheap yet smarter than Claude Sonnet 3.7.
Once model manufacturers go public, the money-burning subsidy wars will definitely stop.
ROI can’t solve everything; SOTA models will never be usable in the coding field.
All projects in this field, since the invention of coding, simply don’t require such consumption.
“A cup of tea, a cigarette, debugging all day” has already become the standard mode for modern programmers.
It’s just that ordinary people have become addicted, entering god mode in the world of code, enjoying the thrill of creating toys.
The fact that Codex users haven’t experienced exponential growth already speaks volumes.
Toys are always toys; the gap between engineering and hobbyists is unimaginably large.
In the future, AI that can make money will only exist in high-value, high-cost industries.
Even if atomic-bomb models are released, they will only make money in niche vertical fields.