Kimi K3 API relay: 1M output tokens for $4.67 on CheapRelay

01 / From an empty account to a $0.09 testing balance

Start with the simplest useful number: a new CheapRelay account receives $0.09 in sign-up credit. That credit works on any model in the price list, so you can test the service before committing to a larger top-up. There is no monthly fee, and the prepaid balance does not expire. That makes the first step measurable rather than theoretical: register, accept the credit, and check which balance you are looking at.

CheapRelay separates ordinary prepaid usage from referral rewards. Your $0.09 sign-up credit belongs to the usable prepaid balance. Referral rewards accumulate in a withdrawable rewards balance instead. Keeping those two balances distinct matters when you plan the first paid call: one pays for API usage, while the other is set aside for an Alipay withdrawal once it reaches the ¥20 minimum.

• ■ Sign-up credit: $0.09, usable on any listed model.
• ■ Billing model: prepaid balance, charged per real token.
• ■ Monthly charge: none.
• ■ Balance expiry: no expiry.

For a frugal test, treat the $0.09 as a small controlled experiment. Pick one model, send one short request, and read the resulting token consumption. If the request uses 10,000 input tokens and 2,000 output tokens on Kimi K3, the example cost is $0.009333 for the input plus $0.00934 for the output, or $0.018673. That fits inside the $0.09 credit, leaving $0.071327 of prepaid balance. The point is not to squeeze the credit into an enormous request; it is to verify the configuration before real usage begins.

02 / How referral rewards become withdrawable money

CheapRelay pays referral rewards as a percentage of what invited users actually pay. A directly invited user contributes 20% of their payment to your withdrawable rewards balance. A user reached through your second level contributes 6%. These percentages apply to payments, not to your own sign-up credit or to money that has not been paid.

Here is the arithmetic. In an example, a directly invited user pays $10 for API usage. Your reward is 20% of $10, which is $2. If a second-level user pays $10, the reward is 6% of $10, which is $0.60. Together, those two example payments add $2.60 to the withdrawable rewards balance. No referral work needs to be converted into a vague “discount”; the site states the share, and the payment supplies the base.

The withdrawal rule is equally direct. The smallest withdrawal is ¥20, and withdrawals are handled through Alipay. An example rewards balance of ¥19.90 remains below that threshold, while a balance of ¥20.00 meets it. The reward percentages are calculated in dollars from paid usage, while the withdrawal threshold is expressed in yuan, so track the two balances separately and avoid treating them as interchangeable.

That separation also prevents a common budgeting mistake. If your prepaid balance is $5 and your withdrawable rewards balance is ¥24, the $5 pays for calls and the ¥24 is available for an Alipay withdrawal once the account is ready. CheapRelay keeps the money used for API consumption separate from money earned through referrals, which makes it easier to check what can be spent and what can be taken out.

03 / Fund the prepaid balance before the first bill

The referral route and the usage route meet at the top-up screen. To make a larger first call, add funds to the prepaid balance. CheapRelay charges for real tokens as they are used, rather than charging a recurring subscription fee. The unused balance remains available because it does not expire.

There is also a refund window. If you top up and then request a refund within 12 days of that top-up, the unused balance is refundable, minus what was spent and the payment fee. An example: after a $20 top-up, you spend $6.50. The remaining $13.50 is the unused portion, but the refund is reduced by the $6.50 already spent and by the applicable payment fee. Plan top-ups around the calls you expect to make; the refund rule is a fallback, not a substitute for sensible budgeting.

Before sending the first substantial request, choose a model whose per-token price matches the job. Kimi K3 lists $0.9333 per 1 million input tokens, $4.67 per 1 million output tokens, and $0.0931 per 1 million cache-read tokens. Gemini 2.5 Flash lists $0.1778 input, $1.42 output, and $0.0181 cache read per 1 million tokens. Kimi K3 output therefore costs $3.25 more per 1 million tokens than Gemini 2.5 Flash output: $4.67 minus $1.42 equals $3.25. That is a useful comparison between two price-list entries, not a claim about an unpublished price elsewhere.

For a mixed request, the arithmetic stays simple. An example using 1 million input tokens and 100,000 output tokens on Kimi K3 costs $0.9333 plus $0.467, which is $1.4003. If 1 million tokens come from the cache instead, the example cache-read charge is $0.0931. Multiply by the actual token counts, add the applicable token categories, and you have a bill you can reproduce with a calculator.

04 / Configure Kimi K3 and make the first billed call

For the first walkthrough, use Kimi K3 as the model id: kimi-k3. Kimi K3 belongs to the OpenAI-compatible group, so point an OpenAI-compatible client at the CheapRelay API base, api.cheapapirelay.com, under /v1. Suitable client examples include Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL. The client and protocol pairing is the part to get right.

Gemini models follow a different path: use the native protocol under /v1beta through Gemini CLI. Claude models require the official Claude Code or Claude Desktop client, version 2.1.285 or newer. Do not route Claude through Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK, cURL, or an OpenAI-compatible configuration. CheapRelay is not an official Anthropic, OpenAI, or Google service; it is a relay with model-specific client requirements.

Once the base URL, path, key, and model id are set, send a short request. The call is billed against the prepaid balance according to the real input, output, and cache-read tokens it consumes. Using the earlier Kimi K3 example of 10,000 input tokens and 2,000 output tokens, the billed amount is $0.018673, taken from the $0.09 sign-up credit. That is the first billed call: one request, one visible token count, one arithmetic result.

Keep the service limits in view while testing. Each key allows 120 calls per minute, and each account allows 8 calls at once. If a request returns HTTP 402, the balance is at or below zero. That status is useful because it gives a direct instruction: add prepaid funds before sending more traffic. It is not a vague failure to troubleshoot.

05 / Check the price-list arithmetic before scaling up

CheapRelay’s selling point is its published price list and the visible gap to official list prices where those prices exist. For Claude and GPT models, the official list prices are shown at 7.2 yuan per US dollar. Other models have no official list price in this comparison, so their relative prices are checked against the listed figures rather than against an invented benchmark.

GPT-6.1 Sol lists $0.1486 per 1 million input tokens and $0.7375 per 1 million output tokens. Its official list prices are $2.00 input and $10.00 output. In an example covering 1 million output tokens, the difference is $10.00 minus $0.7375, or $9.2625. The price list marks that output rate as 93% under official list.

Claude Sonnet 5 lists $0.9764 input and $4.88 output per 1 million tokens, against official list prices of $2.00 and $10.00. The example output difference is $10.00 minus $4.88, or $5.12, and the price list marks the rate as 51% under official list. GPT-6 Astra shows a $0.5639 input rate and $2.82 output rate against official prices of $10.00 and $50.00; the output difference is $47.18, with the price list showing 94% under official list.

That is the whole approach: register with $0.09, earn referral shares of 20% directly and 6% at the second level, withdraw through Alipay from ¥20, top up only the prepaid balance you expect to use, and verify each call against the price list. CheapRelay makes the arithmetic visible, so you can check the invoice instead of trusting a slogan.