Cursor confirms Fable 5 leads every model on CursorBench

Cursor just put Fable 5 back in front of users, and the timing says a lot. The model had apparently been pulled or restricted before, and its return comes with a headline result: Fable 5 now sits at the top of CursorBench, Cursor's internal leaderboard for coding model performance.

That benchmark matters because it's not a marketing number pulled from a paper. It's the yardstick Cursor uses to decide which models actually get recommended inside its own editor, where developers are running real tasks against real codebases all day. Topping it means Fable 5 is, by Cursor's own measure, doing a better job than GPT-4-class and other competing models at the kind of work its users care about.

But the win isn't free. Cursor is upfront that Fable 5 is also the most expensive model per task on the list, meaning every completion, edit, or agentic run costs more than the alternatives. That's the tradeoff nobody gets to skip: raw capability at the top of a leaderboard rarely arrives at bargain pricing, and coding assistants are no exception.

For teams weighing which model to default to, that turns a simple leaderboard placement into an actual cost-benefit decision. Do you pay the premium for the model that wins the benchmark, or do you stick with something cheaper that gets close enough for most day-to-day tasks? Cursor didn't answer that question for anyone. It just handed over the numbers and let the tradeoff sit there in plain view.