Why Cursor gets expensive faster than you use it
The complaint is always the same shape: usage went up a little, the bill went up a lot. Two mechanics produce that gap, and neither one shows up in the token counter people actually look at.
The first is that agent mode re-sends the conversation. Every turn ships the full history back to the model, not the new message. Turn 1 sends your prompt. Turn 20 sends your prompt, nineteen replies, every file read, and every tool result in between. Input does not grow with your typing. It grows with the number of turns.
The second is that the direction that matters costs more per token.
https://aicomp.ai/v1).
Create one free →
What one day of agent work actually costs
Take a realistic day: a few hundred thousand input tokens pulled in by file reads and history re-sends, and a much smaller volume of output. A day of work on this page's workload is 300k input and 60k output tokens — 6M input and 1.2M output across a 20-day month.
Here is that month priced across models, at per-million rates:
| Model | Gateway rate in / out per 1M tokens | One day | 20 days |
|---|---|---|---|
| claude-sonnet-5 | $1.00 / $5.00 | $0.60 | $12.00 |
| claude-opus-5 | $2.5 / $12.5 | $1.50 | $30.00 |
| gpt-5 | $0.625 / $5.00 | $0.49 | $9.75 |
| glm-5.3 | $0.70 / $2.2 | $0.34 | $6.84 |
| deepseek-v4-flash | $0.15 / $0.60 | $0.08 | $1.62 |
One day = 300k input + 60k output on this page's workload. 20 days = 6M input and 1.2M output. Rates checked 2026-10-22 (gateway rates move with upstream promotions), so verify the current number in your dashboard before committing to a budget.
Read the Claude Sonnet 5 row carefully: input carries five times more tokens than output and the two directions cost the same money. That is the whole problem in one line. A pricing page that sorts by input rate will tell you Sonnet is cheap, and on an agent workload it is not.
The spread across the table is roughly 18x from top to bottom. Same workload, same month, same quality of work.
The three things that quietly inflate an agent bill
History re-send. Every turn re-ships the context. A 20-turn session pays for turn 1's tokens twenty times. Cutting turns is worth more than cutting words. One session that ends in 6 turns costs a fraction of the same work spread over 25.
Tool results land in input. Every file read, grep result, and command output becomes input tokens on the next turn. Reading a large file "just in case" is not free, and it is not a one-time cost: it stays in the context for every remaining turn.
Retries are invisible. A failed tool call that the agent retries pays for the full context again. On a long session, a retry near the end is one of the most expensive single events in the session, and it produces nothing.
None of these appear in a per-token rate. That is why comparing rates alone never explains the bill.
Where this leaves you
Two levers, in order of effect: shorten sessions so the context stops compounding, and pick a model by output rate rather than input rate. The first is free and usually bigger. The second is a table lookup away.
FAQ
FAQ
Why does my Cursor bill jump when I barely used it more?
Because agent cost tracks turns, not messages. Adding turns to an existing session re-sends everything already in it, so the marginal cost of turn 20 is far higher than the marginal cost of turn 2. Usage counters that count requests will not show this.
Does a longer context window make it cheaper or more expensive?
More expensive per turn, cheaper per unit of work. A bigger window lets one session finish a task that would otherwise take three sessions, and avoiding the cold start of two extra sessions is usually worth more than the extra tokens per turn.
Is switching to a cheaper model the first thing to try?
It is the second. Sorting by output rate matters (the median model charges 4x more for output), and agent work is output-heavy, but shortening sessions reduces both directions at once and costs nothing.
Do retries really add up?
On long sessions, yes. A retry re-sends the entire accumulated context for zero progress. One retry at turn 18 can cost more than the first five turns combined.