Cursor's Vanishing Price Tag: Why an AI Code Editor Pulled Its Own Cost Data

One line in Cursor's changelog for its new Router does a lot of quiet work: the routed model "can be displayed or hidden," and hidden is the default. When you fire off a coding request now, an intelligent router picks the underlying model for you, and by default you never see which one answered.

That default is the last step of a three-part journey. Cursor has changed how it bills for AI roughly three times in a year, and each change made the real cost of a single request harder to see: first a fixed quota you could count, then volatile token metering, now a router that hides the very model whose rate you're paying.

Before June 2025, Pro cost $20/month and handed you 500 fast premium requests. One prompt to a premium model burned one request. You could count them on one hand at a glance, and staff had earlier anchored the internal figure at roughly $0.04 per request.

The catch was baked in from the start. A heavy multi-file refactor and a one-line tweak both cost "one request," so Cursor absorbed the variance itself. That is exactly why the model couldn't hold: the hardest requests cost far more to serve than the simplest, and a flat quota buries that gap.

On June 16, 2025, Cursor swapped the 500-request quota for a $20 pool of usage credits billed at the underlying model's live API rates, and added a $200/month Ultra plan. Auto mode was pitched as "unlimited," but only Auto was, and that limit wasn't spelled out clearly to users.

People watched their pools evaporate. Secondhand reports described one user going from about $28/month to $500 in three days, and another burning the whole $20 plus $18 in overage in a single day. Predictable monthly spend had turned into a live meter nobody could read in advance.

On July 4, 2025, Cursor admitted the changes "were not communicated clearly," called it "our mistake," and offered refunds for surprise charges racked up between June 16 and July 4.

The new scheme did surface real numbers. Auto billed roughly $0.25 per million cached tokens, $1.25 per million input, and $6 per million output. On paper you could sit down and compute what a request would cost.

In practice you couldn't, because you never know the inputs ahead of time. Cost rides on how much context the agent pulls in and how long the model reasons, so two prompts that look identical diverge sharply. Staff said the hardest requests run an order of magnitude more than simple ones, and credits burn at the same rate as calling the APIs directly.

The Router closes the loop. Cursor pitches it as "frontier-quality results at 60% lower cost," reaching for a cheaper model when the task doesn't need the best one. It runs three modes, Intelligence, Balance, and Cost, and Balance and Intelligence both bill at whatever the routed model's rate happens to be.

So your bill now depends on a model the app selects and, by default, doesn't name. Router is on by default for Teams, and admins can force the model to display, restrict modes, or block specific models, but the individual developer sees a hidden pick. Core features like Tab and Apply always run on Cursor's own models too, so you can't route them to your own API key even when you'd rather own the cost.