I think that tools like this have to operate on a subscription model like Cursor does in order to make any kind of sense for most users. The pay as you go model for agentic code tools makes you responsible for paying for:

* Whatever context the agent decides to pull in.

* However many iterations the model decides to run.

* Any result you get, regardless of how bad it is.

With pay as you go, the tool developer has no incentive to minimize any of these costs—they get paid more if it's inefficient, as long as it's not so inefficient that no one uses it. They don't even need it to be especially popular, they just need some subset of the population to decide that costs don't matter (i.e. those with Silicon Valley salaries).

With Cursor's model of slow and fast queries, they are taking responsibility for ensuring that the agents are as cost efficient as possible. The more efficient the agent the larger their cut. The fewer times that people have to ask a question a second time, the larger their cut. This can incentivize cutting corners, but that somewhat balanced out by the need to keep people renewing their subscription, and on the whole for most users it's better to have a flat subscription price and a company that's optimizing their costs than to have a pay-as-you-go model and a company that has no incentive to improve efficiency.