GLM-5.3 setup: what $0.09 buys you on CheapRelay

01 / What $0.09 of sign-up credit covers on day one

CheapRelay gives every new account $0.09 of sign-up credit, usable on any model. That is small money, so I want to spend it on arithmetic rather than vibes. GLM-5.3 is the model I picked for this first-day walk-through, and the price list shows GLM-5.3 at input $0.4583 per 1,000,000 tokens, output $1.38 per 1,000,000 tokens, and cache read $0.1153 per 1,000,000 tokens. Billing is prepaid balance, pay per real token, with no monthly fee, and the balance does not expire.

Example: if a first day uses 50,000 input tokens and 20,000 output tokens on GLM-5.3, the input side is 50,000 divided by 1,000,000 times $0.4583, which is $0.022915. The output side is 20,000 divided by 1,000,000 times $1.38, which is $0.0276. Adding those two gives $0.050515, so that example day lands under the $0.09 credit, leaving $0.039485 of the credit unused. The arithmetic is the point: CheapRelay charges per real token, so I can see exactly where each cent goes.

For a second example, suppose the same 50,000 input tokens arrive as cache reads instead of fresh input. The cache read line is 50,000 divided by 1,000,000 times $0.1153, which is $0.005765, and with 20,000 output tokens the total is $0.005765 plus $0.0276, which is $0.033365. I am not claiming anything about how often caching happens; I am only showing that the price list gives cache read a lower number than input for GLM-5.3, and the bill follows the token type you actually pay for.

02 / The setup I used for GLM-5.3

GLM is one of the OpenAI-compatible families, so the setup is short. The API base is https://api.cheapapirelay.com, and I point any OpenAI-compatible client under /v1 at it. That list includes Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK or cURL. In the model field I enter the model id exactly: glm-5.3. The limits I work within are 120 calls per minute per key and 8 calls at once per account, and an HTTP 402 tells me the balance is at or below zero.

I keep the client choice deliberately boring, because CheapRelay spells out which client goes with which model. Claude models go only with the official Claude Code and Claude Desktop clients, version 2.1.285 or newer, and never with Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK or cURL, and never OpenAI-compatible. Gemini models use the native protocol only, under /v1beta, through Gemini CLI. GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and the other listed models use the OpenAI-compatible path under /v1. Following that mapping is what keeps the first bill readable.

03 / The before bill: reading the price list before spending

Before I send anything, I write the GLM-5.3 line down: input $0.4583, output $1.38, cache read $0.1153, each per 1,000,000 tokens. Then I compare it with neighbors on the price list, using only the numbers shown. GLM-5.3 input at $0.4583 is lower than Grok 4.6 input at $0.4653, and GLM-5.3 output at $1.38 is lower than Grok 4.6 output at $1.39. Going the other way, DeepSeek V4 Flash shows input $0.2389 and output $0.9556, so both of those numbers sit below the GLM-5.3 numbers. That is the honest framing: GLM-5.3 sits between these two on the price list.

There is no official list price for GLM-5.3 here, so I do not borrow one from another model. For contrast, I can look at the rows that do publish one. GPT-5.6 Sol shows input $0.3056, output $1.71, cache read $0.0264, against official list input $4.00, output $20.00, cache read $0.40; the output here is 91% under list. Claude Sonnet 5 shows input $0.9764, output $4.88, cache read $0.0861, against official list input $2.00, output $10.00, cache read $0.20; the output here is 51% under list. Those gaps are why I check the published numbers instead of assuming a subscription is the only route.

04 / The after bill: one worked day on GLM-5.3

Example: a heavier day runs 200,000 input tokens and 80,000 output tokens on GLM-5.3. Input is 200,000 divided by 1,000,000 times $0.4583, which is $0.09166. Output is 80,000 divided by 1,000,000 times $1.38, which is $0.1104. The day totals $0.20206. Against the $0.09 sign-up credit, that example uses the whole credit and needs $0.11206 more of prepaid balance, which is exactly the kind of gap I want to see on paper before I plan a week.

Example: a cache-heavy day uses 200,000 cache read tokens and 80,000 output tokens. Cache read is 200,000 divided by 1,000,000 times $0.1153, which is $0.02306, and output is $0.1104, for a total of $0.13346. Compared with the fresh-input example at $0.20206, the difference is $0.0686, and that difference is just the two price numbers applied to the same token counts. CheapRelay shows both lines on the price list, so the bill reflects what I sent.

05 / How I top up, and what I can take back

Since billing is prepaid balance with no monthly fee and the balance does not expire, I top up in small amounts and watch the per-token lines. If I change my mind, refunds cover unused balance within 12 days of the top-up, minus what was spent and the payment fee. That policy fits a first-day experiment: I can fund a GLM-5.3 test, read the bill, and stop there without carrying a commitment.

CheapRelay also has a referral line worth knowing: 20% of what a directly invited user pays, and 6% from the second level, paid into a withdrawable rewards balance, with the smallest withdrawal at ¥20 via Alipay. I treat that as arithmetic on other people's spend, not as a reason to change my own model choice. My own choice stays driven by the price list: GLM-5.3 at $0.4583 input, $1.38 output, $0.1153 cache read per 1,000,000 tokens, with the $0.09 sign-up credit covering the first small example above.

06 / What the first bill taught me

Three numbers carried the day for me. First, the GLM-5.3 line is $0.4583 input, $1.38 output, $0.1153 cache read per 1,000,000 tokens. Second, the sign-up credit is $0.09, which the 50,000 input and 20,000 output example at $0.050515 fits inside. Third, the heavier 200,000 input and 80,000 output example at $0.20206 shows what a bigger day costs, and the cache variant at $0.13346 shows the effect of the cache read price. CheapRelay makes those lines visible before and after I spend, which is what I want from a relay.

For the setup itself, nothing exotic: an OpenAI-compatible client under /v1, the API base https://api.cheapapirelay.com, and glm-5.3 in the model field. That is the whole first-day loop, and the bill is the evidence.