Your VS Code Assistant Bill, Line by Line

01 / A $0.35 Example Receipt, Not a Price Promise

The useful question about a VS Code assistant is not simply what one request costs. It is what a finished task costs after context, generated code, cache activity, and follow-up attempts. Start with a receipt you can audit rather than a vague promise of cheap coding.

The example below uses deliberately hypothetical rates, not published CheapRelay prices or measured developer usage. Its purpose is to show the arithmetic. Replace every rate with the applicable rate for your selected model before using the result as a budget.

02 / Connect the Editor Before Counting Tokens

Continue provides one concrete route inside VS Code. Its documentation supports configuring an OpenAI-compatible provider with provider set to openai and an apiBase override. The model entry also needs the corresponding model identifier and API credentials. Use the exact identifier supplied by your service rather than guessing from a display name. ([docs.continue.dev](https://docs.continue.dev/customize/model-providers/top-level/openai?utm_source=openai))

For CheapRelay, the API host is api.cheapapirelay.com, with OpenAI-compatible routes under /v1. That is the versioned base to configure for this workflow; the separate /v1beta route serves Gemini-native requests. CheapRelay offers pay-as-you-go access through one key, without a monthly fee or subscription.

A configurable endpoint is not a guarantee that every assistant feature works with every model. Continue documents model capabilities separately, including explicit capability configuration. Test a short conversation and one small edit before handing over a repository-wide task. ([docs.continue.dev](https://docs.continue.dev/customize/deep-dives/model-capabilities?utm_source=openai))

03 / The Four-Line Token Receipt

Assume one work session produces the following usage across all its requests. These categories are mutually exclusive: cache reads and writes are not counted again as uncached input.

• ■ Uncached input: 120,000 tokens at an assumed $1.00 per million = $0.12.
• ■ Output: 30,000 tokens at an assumed $3.00 per million = $0.09.
• ■ Cache reads: 400,000 tokens at an assumed $0.10 per million = $0.04.
• ■ Cache writes: 80,000 tokens at an assumed $1.25 per million = $0.10.

Total: 630,000 billed tokens and $0.35 in illustrative API usage. Each line follows the same calculation: token count divided by 1,000,000, multiplied by that category's rate. There is no claimed benchmark here, and these rates are not assigned to any actual model.

CheapRelay bills real token usage, including cache reads and writes. Consequently, an honest estimate needs those categories rather than treating caching as free. Also confirm how the selected model's billing defines each category before copying this receipt structure.

04 / Twenty Workdays Turn Cents Into a Budget

If that exact hypothetical session happened once per workday for 20 days, the API usage total would be $7.00. Two identical sessions per day would be $14.00. These are scenario calculations, not predictions about your workload.

Now add a repair attempt using 60,000 uncached input tokens and 10,000 output tokens at the same assumed rates. That adds $0.06 plus $0.03, bringing the session to $0.44. Repeated over 20 days, the total becomes $8.80: $1.80 more than the original scenario.

Keep any separately charged editor software outside this API subtotal. CheapRelay's lack of a subscription does not establish the pricing of whichever extension you choose.

05 / Keep a Receipt for Each Accepted Change

For your first few tasks, record the model identifier, applicable rates, usage categories, total API charge, and whether the resulting change was accepted. Include repair attempts in the same task receipt. That gives you a practical comparison: dollars per accepted change, not dollars per impressive-looking response.

Set a small trial budget before expanding usage. Sign up for CheapRelay, connect your chosen compatible VS Code assistant, and reconcile one narrowly scoped task against its usage records. The goal is not to reproduce this $0.35 example; it is to build a receipt whose numbers are actually yours.