How to use fewer tokens in Cursor: 6 habits that cut usage

Your agent ran for ten minutes, and the usage meter moved more than the task deserved. The short answer: Cursor bills per million tokens, and a long chat makes every new step carry the whole history again. Keep threads short, keep context small, and don't use an expensive model for small edits.

What you are actually billed for#

Cursor's models and pricing page (checked 24 October 2026; the page shows no publication date) lists prices per million tokens, split into four parts: input, cache write, cache read and output. Input is what Cursor sends to the model: your message, the files, the history. Output is what the model writes back. Cache write and cache read are the discounted parts for text that was already sent.

Two more details from the same page matter for habits:

• Long context costs more. Grok 4.7 input above 256k tokens is billed at 2x; some Claude 1M variants are 2x above 200k input. The details are model-specific, so check the page for the model you use.
• Third-party models (Claude, GPT, Gemini) are not in the included weekly pool. On Premium and above they are billed at API list price from prepaid credit; they are not available on Free or Plus.

The habits below follow from how token billing works: every step re-sends what is already in the thread.

Habit 1: start a new chat when the task changes#

An AI model has no memory between requests. Every step of an agent re-sends the thread so far, which is the effect described in why a long AI chat costs more. Step 5 is cheap. Step 150 carries 149 steps of history.

Say you finish a refactor in a thread, then ask the same thread to fix a typo in the README. Nothing in the thread helps with the typo, but all of it gets billed again. Open a new chat, name the file, done.

Cursor's summarization docs (24 October 2026) add one thing worth knowing: restoring a checkpoint reverts files only and does not remove messages from the conversation. If you undo a bad edit and keep chatting, the bad attempt still rides along. When an attempt went wrong, a fresh chat is cleaner than a restore.

Habit 2: point at the file, not at the folder#

A large pasted file or a pile of @-mentions goes into input on every step that follows. Ten mentioned files of 2,000 tokens each add 20,000 tokens to each step, and with a 100-step session that is 2,000,000 input tokens for context the model mostly didn't need.

Mention the one file or function the task touches. If the model needs more, it can ask or read it itself.

Habit 3: stay under the long-context line#

Because long context is billed at a higher rate on some models, a thread that crosses the threshold gets expensive twice: more tokens, and a higher rate on them. If you see a thread growing toward a few hundred thousand tokens, that's the signal to wrap up and start again with a short summary you write yourself: two or three lines on what's done and what's next.

Habit 4: match the model to the job#

Renaming a variable does not need the same model as designing a module. On the models page, the third-party models are billed at API list price, so an expensive model on small edits is the fastest way to burn credit. Use the heavy model for planning and hard bugs; use a cheaper one for routine edits and questions.

Auto has Cost, Balance and Intelligence modes. Individual plans only have Cost mode, which draws from the weekly pool whichever model serves the request. Balance and Intelligence are for Teams and Enterprise with Cursor Router.

Habit 5: ask for short answers when you want short answers#

Output tokens are billed at a higher rate than input tokens on most models (see input vs output tokens). "Show only the changed function" costs less than "rewrite the file" and is easier to review.

Habit 6: move routine chat work to a cheap model on your own key#

For everyday questions and edits in chat, on plans that allow it, you can point Cursor at a custom OpenAI-compatible key. Cursor's models page (24 October 2026) says your own API keys aren't available on Plus; on Teams and Enterprise Cursor adds its own token rate on top of API prices, which the costs below don't include. How to set it up, and which Cursor features work with a custom key, is in How to connect Cursor to CheapTokens. The base URL is https://api.cheaptokens.dev/v1.

Which model? For long sessions with big context, deepseek-ai/DeepSeek-V4-Flash-0731 is the one to try: one price at any time of day, 380K context, up to 32,768 output tokens per answer. For routine edits and quick questions, MiniMaxAI/MiniMax-M2.7 has 180K context and up to 8,192 output tokens per answer, and both its input and output are cheaper than OpenRouter.

What a session costs: before and after#

Let's put numbers on habit 1. Assume an agent session of 200 steps with about 300 output tokens per step, so 60,000 output tokens in total. In one endless thread, history grows and the average step sends around 10,000 input tokens: 200 × 10,000 = 2,000,000 input tokens. With a fresh chat every dozen steps or so, the average drops to around 3,000: 200 × 3,000 = 600,000 input tokens. Same work, same output.

On MiniMax-M2.7, token cost at today's API prices:

• one long thread (2,000,000 in, 60,000 out): $0.109
• short chats (600,000 in, 60,000 out): $0.039

The input shrinks to under a third, and that is the part you control with habits 1 to 3. Here is the long-thread session on every model we have:

These are live prices from the price list, they update when tariffs change. For how input and output split on real tasks, see how many tokens an API request uses and input vs output tokens.

Quick check before you hit send#

1. Did the task change since this chat started? New chat.
2. Are you mentioning more than the files the task touches? Trim.
3. Is this a small edit on an expensive model? Switch.
4. Does the thread feel long? Write two lines of summary and start fresh.

To run chat work on a cheap model, start with the CheapTokens docs and the pricing page. Prices on Cursor's side change, so check its models and pricing page before you plan a month.