$0.1208 per 1M tokens: CheapRelay Codex pay-as-you-go for students

01 / Myth: a student budget needs a fixed recurring payment

A student with a small allowance should not have to commit a fixed amount every month and hope to use it all. CheapRelay uses pay-as-you-go billing: you add a prepaid balance, then pay for the real tokens your requests consume. There is no monthly fee, and the balance does not expire. That makes the cost visible at the point of use rather than hiding it inside a recurring charge.

The price list is measured per 1,000,000 tokens, with separate figures for input, output, and cache reads where applicable. Take Qwen3.7 Plus, whose model id is qwen3.7-plus. Its input price is $0.1208, its output price is $0.3625, and its cache-read price is $0.0361 per 1,000,000 tokens.

Example: one request using 100,000 input tokens and 100,000 output tokens would cost 0.1 multiplied by $0.1208, plus 0.1 multiplied by $0.3625. That is $0.01208 plus $0.03625, or $0.04833. If another 100,000 tokens are read from the cache, add 0.1 multiplied by $0.0361, which is $0.00361, for a total of $0.05194.

That arithmetic is the point. You can calculate a mixed workload before you spend anything. There is no need to buy access in advance, and there is no monthly charge to justify later. For a student whose allowance arrives in small amounts, a prepaid balance that stays there until it is used is a straightforward arrangement.

02 / Myth: signup credit and prepaid money are awkward extras

New accounts receive $0.09 in signup credit, usable on any model. It is not tied to one particular price tier, so a student can try a model without first deciding which one deserves the balance. The credit also arrives before the top-up decision becomes important, which makes it useful for checking the setup.

Example: with $0.09 applied entirely to Qwen3.7 Plus input at $0.1208 per 1,000,000 tokens, the credit covers $0.09 divided by $0.1208, or about 745,033 input tokens. If the same credit is used for output at $0.3625 per 1,000,000 tokens, it covers $0.09 divided by $0.3625, or about 248,275 output tokens. The model choice changes the workload the credit can support, so the price list is worth checking before a big assignment.

The refund rule is also concrete. Unused balance can be refunded within 12 days of the top-up, minus what was spent and the payment fee. That is not a vague promise about account activity; it defines the window and the deductions.

CheapRelay also has a referral program. A directly invited user generates rewards equal to 20% of what they pay, while the second level contributes 6%. Those rewards go into a separate withdrawable rewards balance. The smallest withdrawal is ¥20, through Alipay. For students sharing a tool with classmates, the referral percentage provides another way to estimate the value of the program without adding a subscription cost.

03 / Myth: every AI client should be able to call every model

Here is where a little arithmetic saves a lot of debugging. CheapRelay does not treat every model as the same protocol, and using the wrong client can lead to a failed connection that has nothing to do with your balance. Check the model family and the client before configuring anything.

Claude models, including Claude Haiku 4.5, Claude Sonnet 5.5, Claude Sonnet 5, Claude Opus 5.5, and Claude Opus 5, use the official Claude Code and Claude Desktop clients, version 2.1.285 or newer. They do not run through Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK, or cURL, and they are not accessed through an OpenAI-compatible endpoint.

Gemini models follow a different route. Gemini 2.5 Flash, Gemini 3.8 Flash, and Gemini 3.1 Pro use the native protocol under /v1beta through Gemini CLI. They are not OpenAI-compatible in this setup.

For GPT models, including access through Codex, and for DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax, and the other models on the price list, use an OpenAI-compatible client under /v1. Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, and cURL are examples of suitable clients. The base address is https://api.cheapapirelay.com.

Example: if your coursework uses Gemini CLI, configure Gemini 2.5 Flash with the model id gemini-2.5-flash and the native /v1beta path. If the same coursework uses an OpenAI-compatible client, Qwen3.7 Plus uses qwen3.7-plus under /v1. The model id should be copied exactly, because a small naming error can look like a service problem when it is only a configuration mismatch.

04 / Myth: the official list price and the CheapRelay price tell the same story

The price list includes official list prices for the Claude and GPT models, so the gap can be shown rather than merely described. Start with GPT-6.1 Sol, whose model id is gpt-6.1-sol. Its input price is $0.1486 and its output price is $0.7375 per 1,000,000 tokens. The official list prices are $2.00 for input and $10.00 for output. Its cache-read price is $0.0347, compared with an official cache-read price of $0.10.

Example: 1,000,000 input tokens plus 1,000,000 output tokens cost $0.1486 plus $0.7375, or $0.8861 on CheapRelay. At the official list prices, the same token volume would cost $2.00 plus $10.00, or $12.00. The difference is $12.00 minus $0.8861, which is $11.1139. For output alone, $0.7375 compared with $10.00 is 93% under the official list price.

Claude shows a similar arithmetic pattern without borrowing prices from other families. Claude Sonnet 5.5, with model id claude-sonnet-5-5, costs $0.8903 for input and $4.45 for output per 1,000,000 tokens. Its official list prices are $2.00 and $10.00, and its cache-read price is $0.0778 against an official $0.10.

Example: 1,000,000 input tokens plus 1,000,000 output tokens cost $0.8903 plus $4.45, or $5.3403. The official list total is $12.00, so the difference is $6.6597. Its $4.45 output price is 56% under the $10.00 official output price.

GPT-6 Astra, model id gpt-6-astra, costs $0.5639 for input and $2.82 for output, with a cache-read price of $0.0569. The official figures are $10.00, $50.00, and $1.00. Example: 1,000,000 input and 1,000,000 output tokens cost $0.5639 plus $2.82, or $3.3839, compared with an official total of $60.00. Output at $2.82 is 94% under the $50.00 official output price.

For models without an official list price on this page, compare their CheapRelay figures with each other. Example: Qwen3.7 Plus costs $0.1208 for input and $0.3625 for output, while MiMo V2.6 Pro, model id MiMo-V2.6-Pro, costs $0.1736 and $0.3472. One million of each token type costs $0.4833 for Qwen3.7 Plus and $0.5208 for MiMo V2.6 Pro, a difference of $0.0375. MiMo V2.6 Pro is also $0.002778 for cache reads, compared with $0.0361 for Qwen3.7 Plus.

05 / Myth: a small balance is too fragile for coursework

Balance limits are easier to plan for when they are stated as numbers. 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 is a clear signal to top up rather than a vague failure to diagnose.

Example: a student using GPT-6.1 Sol for a session with 500,000 input tokens and 250,000 output tokens would pay 0.5 multiplied by $0.1486, plus 0.25 multiplied by $0.7375. The input portion is $0.0743 and the output portion is $0.184375, for a total of $0.258675. A $0.09 signup credit would not cover that whole session, but it would cover $0.09 of it, and the remaining $0.168675 can be checked against the prepaid balance.

That is the practical value of pay-as-you-go for a student: the budget follows actual use. Start with the $0.09 credit, choose a client that matches the model, and check the price list for the token mix in your assignment. CheapRelay makes the next purchase easier to calculate because there is no monthly fee, the balance does not expire, and every charge is tied to real token consumption.