¥0.60 Credit: A 95-Request Test Receipt

01 / Sixty fen buys a checkout test, not a production trial

A ¥0.60 sign-up credit is sixty fen, or six-tenths of one yuan. Treat it as a small integration budget: enough to check authentication, inspect responses, validate structured output and see whether your application handles errors. It is not a meaningful allowance for unrestricted coding sessions or long-running agents.

CheapRelay is a low-priced, pay-as-you-go AI API relay with no monthly fee or subscription. Its billing follows real token usage, including cache reads and writes. That makes a small credit useful, but only if you set a request budget before pressing run.

02 / The receipt assumes ¥1 input and ¥4 output per million tokens

The following calculation uses illustrative rates of ¥1 per million input tokens and ¥4 per million output tokens. These are arithmetic assumptions, not a CheapRelay price quote or a claim about any particular model. Check the current rates for your selected model before using this receipt as a spending plan.

At those assumed rates, a request containing 2,000 input tokens and 1,000 output tokens costs ¥0.006: ¥0.002 for input plus ¥0.004 for output. That is 0.6 fen per request. Actual output length can change the bill, so the counts below are planning targets, not guaranteed request allowances.

• ■ Authentication and basic responses: 30 requests × 2,000 input and 1,000 output tokens = ¥0.18.
• ■ Structured-output checks: 40 requests × 1,000 input and 500 output tokens = ¥0.12.
• ■ Application failure-path fixtures: 15 requests × 4,000 input and 1,000 output tokens = ¥0.12.
• ■ Longer-history checks: 10 requests × 10,000 input and 500 output tokens = ¥0.12.
• ■ Unspent reserve for retries or output overruns: ¥0.06.

Total planned usage: 95 requests, 180,000 input tokens and 90,000 output tokens. Input costs ¥0.18; output costs ¥0.36. The planned bill is ¥0.54, leaving six fen untouched. This example assumes ordinary uncached input and does not claim any cache savings.

03 / Give each of the 95 requests a specific job

Use the first 30 requests to verify the connection from your actual application, not just a playground. Check a short response, streaming if your application uses it, response parsing and usage reporting. Stop repeating a test once it passes; identical greetings provide little additional evidence.

Spend the next 40 on structured output. Prepare small fixtures with required fields, empty values, escaped quotation marks and unexpected user text. Validate the results in code. A response that looks like JSON in a terminal is not enough if your parser rejects it.

The 15 failure-path fixtures should test your application's handling of missing information, contradictory instructions and unusable model output. They are not a reason to send deliberately invalid API calls: rejected requests do not reliably exercise the same billing or response path. Simulate transport failures locally where possible.

Use the final ten requests to send a longer conversation or document, then ask for a short answer grounded in it. This checks whether your application preserves context and passes the intended messages. Ten examples cannot establish broad model quality or dependable long-context recall.

04 / Keep setup mistakes outside the paid receipt

CheapRelay's API base is https://api.cheapapirelay.com, with OpenAI-compatible routes under /v1 and Gemini-native routes under /v1beta. Match your client's expected route format; avoid accidentally appending /v1 twice. Store the key outside source control and confirm configuration before launching a batch.

Set an output cap, disable automatic retries during the first checks and inspect usage after each small batch. Repeated conversation history adds input tokens. Tool results can add more. Cache reads and writes are billable categories, so do not treat cached context as free or apply this simple receipt unchanged to a cached workflow.

05 / The useful result is a working integration and a measured bill

Sign up for CheapRelay to use the sixty-fen credit as a controlled integration check, then replace these assumed rates with current pricing. If your chosen rates are twice this example's rates, the same token workload costs ¥1.08 and exceeds the credit. The win is not a promised number of conversations; it is finding configuration and parsing mistakes before funding a larger workload.