Cursor Made the Price Tag Disappear

There’s a strange moment that hits every developer using AI coding tools. You’ve been shipping all month, the agent has been humming along, and then you realize you have absolutely no idea what you’ve spent. Reports have been circulating that Cursor removed cost figures from its usage page and trimmed the related fields from its CSV export. That uneasy feeling is back.

One caveat up front: this topic is thin on community data. Over the past 30 days there’s been essentially no substantial discussion thread to point to. So treat this less as live coverage of a specific incident and more as a look at why AI coding tool billing keeps drifting toward opacity.

A number is not just a number

The cost display on a usage page isn’t a UI element. It’s the window through which you verify the deal: you pay for what you use.

Traditional SaaS made this easy. Twenty dollars a month means twenty dollars a month. AI tools break that. Same $20, and some months you cruise, other months you hit the wall on day 20. The reason is simple — underneath, everything is priced in tokens.

Remove the cost display and users lose two things at once. You can’t see how much you’ve spent, and you can’t see what drove the spending. The second loss is the one that actually hurts.

Token billing is genuinely hard

The economics of an AI coding tool are messier than most people assume, because the cost of a single request isn’t fixed.

Take the same prompt — refactor this function. If your context pulls in 3 files, that’s one price. If it pulls in 30, your input tokens jump by an order of magnitude. Now add an agent that reads files on its own, runs searches, and retries when something fails. You made one request; behind the scenes the tool called the model a dozen times.

Cache discounts add another variable. Reuse the same context and it gets cheap. Churn through files and it gets expensive. Switch models and the per-token rate changes entirely. All of this happens silently.

Which is exactly why a cost dashboard was never a luxury. It was the only place a user could inspect that arithmetic.

The CSV export matters more than the dashboard

Losing the on-screen number stings. Losing the CSV export is the bigger practical hit.

Reading a figure on a screen and pulling raw data into a spreadsheet are entirely different capabilities. With a CSV, a team can aggregate spend per person, trace which projects are eating tokens, and produce something finance will actually accept. That last one matters more than engineers like to admit.

In enterprise settings it’s also a negotiating asset. An org that can say “our team averages this much per month” from its own records sits in a different chair at the contract table than one squinting at a vendor-supplied summary. Lose the raw data and you lose the ability to verify anything.

Why this keeps happening

It’s hard to call this malice. The structural pressure explains more.

AI tool companies are in a genuinely awkward spot. Inference costs are real money going out the door, while users want flat-rate pricing. It’s routine for a single power user to burn several times their subscription fee. So vendors introduce abstractions: request counts, credits, fast requests. These act as a buffer between actual cost and the number you see.

The thicker that buffer gets, the less control you have over your own consumption. Abstraction is convenient, and it quietly removes your ability to audit.

There’s one more piece. When real costs are visible, users start doing math. Wouldn’t it be cheaper to just hit the API directly? The moment that question surfaces, the vendor’s margin structure is on the table. Hiding cost data is commercially rational for exactly this reason.

What you can actually do

There are moves available.

The most reliable one: keep your own logs. If you can’t trust the vendor’s dashboard, start recording billing statements alongside your actual workload. It beats getting surprised by an invoice at month end.

Keep escape routes open. Direct API calls, open-weight models running locally, a competing tool you know how to use. Full lock-in means that when pricing shifts, you have no choice but to follow.

And speak up. Developer tooling still turns user feedback into product changes relatively fast. Enough pressure for transparency does sometimes get features restored.

The takeaway

One missing number looks like a small change. The direction is the problem. As AI tools move into the center of how we work, how transparent a vendor is becomes a selection criterion on par with the technical specs.

Do you know exactly what you spent on AI coding tools this month? And do you still have a way to find out? If the answer doesn’t come quickly, that’s already the signal.