Cursor API cost calculator
If you run Cursor with your own API key, your bill is pure token spend. Here is the monthly number And what the same volume costs through a volume route.
What a month of Cursor actually looks like
The calculator starts at 180M input and 18M output a month, roughly a developer who works in agent mode most of the day. At that volume:
| Model | Per 1M in / out | This month |
|---|---|---|
| Claude Opus 5 | $2.5 / $12.5 | $675.00 |
| Claude Sonnet 5 | $1.00 / $5.00 | $270.00 |
| Claude Haiku 4.5 | $0.50 / $2.5 | $135.00 |
| GPT-5 | $0.625 / $5.00 | $202.5 |
| GPT-5 mini | $0.125 / $1.00 | $40.5 |
| DeepSeek V4 Flash | $0.22 / $0.66 | $51.48 |
Change the two numbers and you have your own figure. If you do not know your token counts yet, where a Cursor bill comes from walks through how to get them.
Why the estimate is usually too low
Cursor in agent mode does not send one message per request. It sends the accumulated thread (the files it read, the tool results, the earlier attempts) on every turn. A session that felt like twenty short messages can easily have moved several million tokens, so the cost grows with thread length, not with how much you typed.
- Long threads are the multiplier. Restarting a session resets the context and is the single cheapest habit here.
- Agent runs dominate, completions do not. Inline edits and tab completions are cheap; judge the bill by agent turns.
- Input decides the number. Cursor traffic is far heavier on input than output, so the input rate is the one to compare.
Cursor with your own key
Cursor lets you supply an Anthropic or OpenAI key instead of using their pooled quota. When you do, you pay raw token cost, which means the per-token rate is the only thing that matters. See the Cursor setup guide.
When the number comes back too high
Three levers, in the order they move the total:
- Shorter threads. Costs nothing and cuts the multiplier directly.
- A cheaper default model. Most agent turns do not need the largest one On this list, Opus-class input costs 5× Haiku-class input.
- A lower per-token rate. Same models, same SDK; the base URL decides who bills you. How that works, then the two-line Cursor change.
Figures are estimates based on list prices verified 2026-11-02. Actual spend depends on your cache hit rate, batch usage and long-context tiering.