Cursor AI pricing: the subscription vs paying per token on your own key

The query "cursor ai pricing" is really two questions: what Cursor itself costs, and what tokens cost if you plug your own key into it. A subscription and tokens are separate line items, and you cannot merge them into one number without doing the math. Below we explain how billing is structured, give a formula, and calculate a week of agent work on several models at Zukko prices as of October 18, 2026.

Cursor AI pricing: what it consists of

Cursor has pricing plans. Their contents and prices change, so look up current figures on cursor.com/pricing↗; we do not quote them here so as not to mislead you. What matters is what exactly the subscription covers and what it does not:

• The subscription gives you the editor itself plus an included amount of usage under your plan's rules. How that amount is counted and what happens when it runs out is described in Cursor's documentation.
• Your own key (BYOK) is connected in the model settings. According to Cursor's documentation, on individual plans the key's provider bills those requests, not Cursor, and they do not draw from the included usage. On team and enterprise plans, per the same documentation, Cursor additionally charges its own per-token rate (the Cursor Token Rate) on such requests; check the amount with Cursor.
• Tab autocomplete keeps running on Cursor's built-in models even if chat and the agent go through your key. A custom key works only with chat models.

The practical conclusion: your own key does not cancel the subscription entirely. It removes the part you spend on agents and chat with third-party models and replaces it with payment for the tokens you actually use.

One caveat about plans. According to Zukko's documentation, overriding the provider requires a paid Cursor plan: it is unavailable on the free one. But Cursor's plan lineup has changed over time, and not every paid plan necessarily includes the ability to use your own keys. Check the plans table before you pay.

The formula: what a request costs on your own key

For a Zukko key, the cost of a single request is:

cost = input × input_price / 1,000,000 + output × output_price / 1,000,000

If the model has a cache, cache reads and writes are added at their own catalog prices. But there is a Cursor-specific nuance. Cursor connects to Zukko as an OpenAI-compatible client (base URL https://api.zukko.pro/v1, endpoint /v1/chat/completions). In that format, for Claude models the entire input is billed at the input price: the gateway cannot see how much came from cache. For GPT models, cache reads are visible in the OpenAI format too. So in the calculations below the cache is ignored for Claude, and both variants are shown for GPT.

Prices per 1M tokens as of October 18, 2026:

The strength index is the Intelligence Index from Artificial Analysis; we cite no other quality ratings. For the current list and prices, call GET https://api.zukko.pro/v1/models or open the "Models" page in the dashboard.

Example: a week of work with the Cursor agent

Take a hypothetical profile. It is an assumption for the calculation, not a measurement: you will see your own numbers in the dashboard's "Usage" section.

• 5 working days, 40 agent requests a day: 200 requests a week;
• on average 30,000 input tokens per request (chat history, open files, rules, tool descriptions);
• on average 1,500 output tokens.

Weekly total: input 200 × 30,000 = 6,000,000 tokens, output 200 × 1,500 = 300,000 tokens.

claude-sonnet-5-5: 6 × $0.4 + 0.3 × $2 = $2.40 + $0.60 = $3.00. The same tokens at official prices: 6 × $2 + 0.3 × $10 = $15.00.

claude-opus-5-5: 6 × $0.8 + 0.3 × $4 = $4.80 + $1.20 = $6.00. Official price: 6 × $4 + 0.3 × $20 = $30.00.

gpt-6.1-sol: 6 × $0.27 + 0.3 × $1.32 = $1.62 + $0.396 = $2.016. If 70% of the input is read from cache (an assumption), that is 1.8M fresh tokens and 4.2M cached: 1.8 × $0.27 + 4.2 × $0.013 + 0.3 × $1.32 = $0.486 + $0.0546 + $0.396 = $0.9366.

glm-5.3-flash: 6 × $0.04 + 0.3 × $0.04 = $0.24 + $0.012 = $0.252.

deepseek-v4.1-flash: 6 × $0.013 + 0.3 × $0.038 = $0.078 + $0.0114 = $0.0894.

Over a month (4 weeks), claude-sonnet-5-5 comes to $12.00, claude-opus-5-5 to $24.00 and deepseek-v4.1-flash to about $0.36. The spread between models is tens of times, so on your own key the model you pick decides what you pay.

To compare with the subscription, take your plan's current price from cursor.com/pricing↗ and add the token spend from the formula above. Whether it comes out cheaper depends on how much you work with the agent.

What inflates spending the most

1. Long chat history. Every new request sends the model the whole accumulated conversation. Start a new chat for each task.
2. Many attached files and folders. Every @ mention adds tokens to all following requests.
3. Large rules and MCP tools. They go into the context of every request. More on rules in the article how to write Cursor rules↗.
4. A model that is too strong for a simple task. There is no need to hand a variable rename to Opus.

A practical scheme: everyday edits on an inexpensive model, a complex multi-file task on claude-opus-5-5. Models are switched right in Cursor's model list.

How to connect your own key

In short: in Cursor open Settings → Models, enable your own OpenAI key, paste fetch_..., enable the Base URL override and enter https://api.zukko.pro/v1, then add a model name and click Verify. Details and a table of alternative model names are in the Cursor documentation↗. For spending control, see the key limits page↗: give the key a daily and monthly ceiling in dollars so the agent cannot run you into the red.

The balance is topped up from $1: via SBP (the payment page adds an 8.5% fee on top, and the balance receives the dollar amount you requested), with cryptocurrency, or with a MasterCard card (fee of about 13%, amounts $2 to $594). For example, to get $10 on the balance via SBP, the payment is $10 × 1.085 = $10.85 in rubles at the rate on the day of payment.

FAQ

Does your own key fully replace a Cursor subscription?

No. Autocomplete stays on Cursor's built-in models, and the ability to add keys depends on the plan. Your own key replaces the part you would spend on chat and the agent with third-party models.

Which is cheaper: the subscription or a key?

It depends on intensity. With light use, tokens cost a few dollars a month; with heavy work on the strongest model the sums grow. Calculate with the formula on your own profile and compare it with the current plan price.

Does Cursor see my requests when I use my own key?

According to Cursor's documentation, all requests go through its servers to build the final prompt, so the gateway address must be reachable from the internet (local addresses will not work). Cursor's zero data retention policy does not apply to requests made with your own key: the provider's rules apply.

Why is there no cache discount for Claude in the calculation?

Cursor uses the OpenAI format, and in it the gateway cannot see how much input came from Claude's cache. The Claude cache discount comes with the Anthropic format (Claude Code, the Anthropic SDK); see the Claude Code vs Cursor comparison↗.

To check the calculation in practice, create a key in the Zukko dashboard↗, connect it in Cursor, and look at the real spending under "Usage" after a day.