Most AI coding tools now bill on a model the industry borrowed from cloud infrastructure: a flat subscription that includes some allowance, then metered usage on top, charged in arrears at the underlying model's API rate.
That is a fundamentally different cost model from a subscription, and most of us evaluate it as though it were the same thing. I paid for Cursor for eleven months across 2025 and then cancelled, and the invoices make the difference concrete in a way the pricing page does not.
The arithmetic
Eleven months, $510.70 total. Of that, $340 was subscription and $170.70 was on-demand usage — a third of the bill, 33.4%, sitting on top of a plan advertised at $20.
The distribution is the interesting part. It was not evenly spread:
| Month | Billed | Share that was overage |
|---|---|---|
| October 2025 | $159.22 | 49.8% |
| November 2025 | $110.00 | 45.5% |
| (11-month total) | $510.70 | 33.4% |
Two months where metered usage was roughly half the invoice. On a plan whose headline number is $20.
Nothing was wrong with the billing. Usage beyond the included allowance is charged at the model's API rate, and on individual plans that is the model's rate — the discounted token rate is a Teams and Enterprise feature. The rate was disclosed. What was missing was any intuition for how fast a heavy month accumulates.
Why the intuition fails
The failure mode is not price, it is variance. A subscription has a known worst case. Subscription-plus-metering has a worst case bounded only by your own throughput, and throughput on an agentic tool is exactly what spikes when work gets hard — the weeks you lean on it most are the weeks it costs most. Cost correlates with difficulty, which is precisely backwards from how you would budget it.
I upgraded to the $60 tier at one point expecting headroom. The included allowance ran out in about a week.
If you want a mental model: treat the subscription as a floor, not a price. The real number is floor + (throughput × rate), and you do not know your throughput until you have a few months of history. Which means for the first quarter you are, in effect, running an uninstrumented cost centre.
The control that exists and nobody sets
Cursor has a spend limit. It is in the billing settings. It has been there the whole time.
Set it on day one — before the first heavy week, not after the first surprising invoice. It converts an unbounded model back into a bounded one, which is the only change that makes the tool's cost predictable enough to defend in a budget conversation.
This generalises past Cursor. Any tool billing subscription-plus-metering should be onboarded the same way:
- Find the spend cap before you find the features. If there is no cap, that is a finding.
- Instrument the first month — you are collecting a throughput baseline, not just doing work.
- Budget the floor plus a variance band, not the sticker price.
- Re-check when you change tier. A bigger plan changes the allowance, not the billing model, and a bigger allowance can encourage exactly the usage that blows through it.
What I concluded
I cancelled — not because it was bad. It was genuinely good, and for a lot of people $20 flat plus modest metering is fair value for what it does. I cancelled because my usage pattern made the metered layer the dominant term, and at that point I would rather pay a predictable number.
That is a judgement about a usage profile, not a verdict on the product. Yours may differ, and the honest way to find out is to instrument it rather than estimate it.
The full breakdown — what it does well, where it struggles, and the thing I got wrong about its billing for months — is in my Cursor review, written from the invoices rather than from the marketing page.
Figures are from my own invoices across 2025. Published tier names are Pro $20, Pro Plus $60, Ultra $200 — I have deliberately not quoted current per-tier allowances, because those have changed repeatedly and any number I cite here would date badly. The billing model is the durable part; the inclusions are not.
Top comments (0)