What does ¥10 buy on CheapRelay API relay with GLM-5.3?

01 / The ¥10 budget behind every CheapRelay bill

Ten yuan sounds small until you put it against a real token bill. CheapRelay charges per real token, with no monthly fee, and the balance does not expire. For this worked example, I am converting the ¥10 budget at the rate used for the official Claude and GPT list prices: 7.2 yuan per US dollar. That gives ¥10 ÷ 7.2 = $1.3889 of API spending. I will round the working figure to $1.3889, then show how many tokens that buys on selected models from the CheapRelay price list.

Every number below is an example, not a promise about how many tokens your own request will consume. Input, output, and cache reads are billed separately. The useful question is therefore not simply “Which model is popular?” It is “What can this particular workload cost on this particular bill?”

CheapRelay also gives every new account $0.09 in sign-up credit, usable on any model. That credit sits alongside the prepaid balance. You pay for the tokens you use, and you do not need to keep paying a subscription fee while a balance sits unused.

02 / Before the bill: turning a workload into token arithmetic

For a repeatable example, let us use a workload containing 1,000,000 input tokens and 1,680,000 output tokens. On GPT-6.1 Sol, the CheapRelay price list shows $0.1486 per 1,000,000 input tokens and $0.7375 per 1,000,000 output tokens. The arithmetic is:

• ■ Input: 1 × $0.1486 = $0.1486.
• ■ Output: 1.68 × $0.7375 = $1.2390.
• ■ Total: $0.1486 + $1.2390 = $1.3876.
• ■ Converted at 7.2 yuan per dollar: $1.3876 × 7.2 = ¥9.99072.

So the “after” CheapRelay bill for this example workload is ¥9.99072. The remaining balance is $1.3889 − $1.3876 = $0.0013, which converts to ¥0.00936.

Now the “before” bill. GPT-6.1 Sol has official list prices of $2.00 for 1,000,000 input tokens and $10.00 for 1,000,000 output tokens. The same workload therefore costs 1 × $2.00 = $2.00 for input, plus 1.68 × $10.00 = $16.80 for output. That is $18.80, or $18.80 × 7.2 = ¥135.36.

The CheapRelay example is ¥9.99072 instead of ¥135.36. The published price-list gap for GPT-6.1 Sol output is 93% under the official list price. This is the sort of arithmetic worth doing before choosing an API relay: put the same token counts on both bills, then read the totals.

03 / After the bill: what ¥10 buys across several model families

GLM-5.3 is a useful starting point because its price list includes $0.4583 per 1,000,000 input tokens, $1.38 per 1,000,000 output tokens, and $0.1153 per 1,000,000 cache-read tokens. For an example workload of 1,000,000 input tokens and 674,000 output tokens, the CheapRelay bill is 1 × $0.4583 = $0.4583 for input, plus 0.674 × $1.38 = $0.93012 for output. The total is $1.38842, or ¥9.996624. In other words, this example buys 1,000,000 input tokens plus 674,000 output tokens on GLM-5.3 for just under ¥10.

Qwen3.7 Plus has a different balance. Its listed prices are $0.1208 input, $0.3625 output, and $0.0361 for cache reads per 1,000,000 tokens. In an example with 1,000,000 input tokens and 3,480,000 output tokens, the total is $0.1208 + 3.48 × $0.3625 = $1.3823, which is ¥9.95256. MiniMax M3 lists $0.1236 input and $0.4917 output per 1,000,000 tokens. An example with 1,000,000 input and 2,570,000 output tokens costs $0.1236 + 2.57 × $0.4917 = $1.387269, or ¥9.9883368.

Gemini 2.5 Flash lists $0.1778 input and $1.42 output per 1,000,000 tokens. For an example of 1,000,000 input tokens and 851,900 output tokens, the calculation is $0.1778 + 0.8519 × $1.42 = $1.387498, or ¥9.9899856. Gemini models use the native protocol under /v1beta through Gemini CLI, not an OpenAI-compatible client.

For Claude Sonnet 5, the price list shows $0.9764 input and $4.88 output per 1,000,000 tokens. An example containing 1,000,000 input tokens and 84,500 output tokens costs $0.9764 + 0.0845 × $4.88 = $1.38876, or ¥9.999072. The same example at the official Sonnet 5 list prices would be 1 × $2.00 + 0.0845 × $10.00 = $2.845, which converts to ¥20.484. The published output gap is 51% under list.

04 / Choosing a model without borrowing someone else’s price

Official list prices are shown for the Claude and GPT models on the price list, so those comparisons are straightforward. For models without an official figure here, the honest method is to compare CheapRelay’s own rows directly. Claude Opus 5 lists $1.66 input and $8.28 output per 1,000,000 tokens, while Claude Opus 5.5 lists $1.32 input and $6.24 output. For an identical example workload, the lower listed figures are the ones on Opus 5.5.

Claude Sonnet 5.5 lists $0.8903 input and $4.45 output, compared with $0.9764 and $4.88 for Sonnet 5. Kimi K3 lists $0.9333 input and $4.67 output. Gemini 3.1 Pro lists $1.12 input and $6.49 output, while Gemini 3.8 Flash lists $0.95 input and $4.28 output. These are direct CheapRelay price-list comparisons, without assigning an official list price to either Gemini model.

Other rows help fill out the budget picture. GPT-5.6 Sol lists $0.3056 input and $1.71 output, with output 91% under its official list. GPT-6 Astra lists $0.5639 input and $2.82 output, with output 94% under list. Claude Haiku 4.5 lists $0.2917 input and $1.40 output, with output 72% under list. Claude Opus 5.5 output is 69% under list, and Claude Opus 5 output is 67% under list.

For OpenAI-compatible routes, models including GPT, DeepSeek, Kimi, GLM, Qwen, Grok, and MiniMax can be used with clients such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL under /v1. Claude is different: use the official Claude Code or Claude Desktop client, version 2.1.285 or newer. That distinction keeps the bill simple because the client and protocol match the model.

05 / What happens after the balance reaches zero

CheapRelay is prepaid and pay-as-you-go. There is no monthly fee, and unused balance does not expire. Each key is limited to 120 calls per minute, with 8 calls at once per account. If a request returns HTTP 402, the balance is at or below zero. That gives you a clear signal to top up rather than guess whether a request failed because of the model, the client, or the account balance.

There is also a refund rule for unused balance within 12 days of a top-up: the refund subtracts what was spent and the payment fee. Referrals can add to the value picture as well. A directly invited user generates a 20% reward based on what they pay, while the second level generates 6%. Those rewards enter a withdrawable rewards balance, with the smallest withdrawal set at ¥20 through Alipay.

The final arithmetic is simple. A ¥10 prepaid balance is $1.3889 at the list comparison rate. Depending on the model and the input-output mix, that can cover the worked Qwen3.7 Plus, MiniMax M3, GLM-5.3, Gemini 2.5 Flash, GPT-6.1 Sol, or Claude Sonnet 5 workloads shown here. CheapRelay makes the unit prices visible, lets you calculate before committing, and charges for the real tokens your work actually uses.