Claude Code via an API proxy costs $0.7375 per 1M output tokens

01 / What CheapRelay charges for Claude Code, Codex and Gemini CLI

Model prices are easy to quote and hard to compare. CheapRelay puts the arithmetic in one place: the price list shows token prices per 1,000,000 tokens, while official list prices are provided for the Claude and GPT models. You can therefore compare the same workload without converting vague subscription allowances into token estimates.

For a simple output-heavy example, suppose a job generates 1,000,000 output tokens and does not bill input tokens. GPT-6.1 Sol through Codex costs $0.7375 at CheapRelay. Its official list output price is $10.00, so the difference is $10.00 minus $0.7375, or $9.2625. That makes the CheapRelay output price 93% under the official list price.

Claude models use the same visible comparison. Claude Sonnet 5 costs $4.88 per 1,000,000 output tokens at CheapRelay, against an official list price of $10.00. The saving on that example is $5.12 per 1,000,000 output tokens, and the output price is 51% under list. Claude Opus 5 costs $8.28 on the CheapRelay price list versus $25.00 officially, a difference of $16.72 and a 67% lower output price.

Gemini models do not have an official list price in this table, so the honest comparison is between Gemini rows. Gemini 2.5 Flash costs $0.1778 per 1,000,000 input tokens and $1.42 per 1,000,000 output tokens. Gemini 3.8 Flash costs $0.95 and $4.28 for the same units. For an example using 1,000,000 input plus 1,000,000 output tokens, Flash 2.5 costs $0.1778 plus $1.42, or $1.5978. Flash 3.8 costs $0.95 plus $4.28, or $5.23. The difference is $3.6322 for that example workload.

02 / Choose the client that matches the model

CheapRelay is not a single protocol wearing three badges. The model determines the client and endpoint, which prevents a frustrating afternoon of incompatible settings.

Claude models are available only through the official Claude Code and Claude Desktop clients, version 2.1.285 or newer. They are not configured through Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK or cURL, and they are not exposed through an OpenAI-compatible route. In other words, if your daily tool is Claude Code, the supported path is Claude Code.

Gemini models use the native protocol under /v1beta through Gemini CLI. They are not OpenAI-compatible models. That distinction matters because copying a generic OpenAI configuration into Gemini CLI does not turn a native Gemini route into an OpenAI-compatible one.

GPT models, including GPT models used through Codex, follow the OpenAI-compatible path under /v1. DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and the other compatible models use that same route. Supported clients include Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK and cURL. The result is a practical split: Claude Code for Claude, Gemini CLI for Gemini, and Codex or another compatible client for GPT and the compatible model family.

The base address is https://api.cheapapirelay.com. When the client asks for a model, enter the model id from the price list rather than improvising a display name. Examples include qwen3.7-plus, gpt-6.1-sol, gemini-2.5-flash, claude-haiku-4-5-20251001 and deepseek-v4-pro.

03 / Build the bill before you build the workflow

A prepaid balance makes the bill inspectable. You add funds, calls consume balance according to the real tokens used, and there is no monthly fee. The balance does not expire, so unused funds are not tied to a renewal date.

Here is a mixed example. Assume one request bills 1,000,000 input tokens, 200,000 output tokens and 1,000,000 cache-read tokens. With Qwen3.7 Plus, the CheapRelay cost is $0.1208 plus $0.3625 multiplied by 0.2, plus $0.0361. The output portion is $0.0725, so the example total is $0.2294.

Now compare that with MiMo V2.6 Pro on the same example. Its input price is $0.1736, its output price is $0.3472 and its cache-read price is $0.002778. The example total is $0.1736 plus $0.06944 plus $0.002778, or $0.245818. Qwen3.7 Plus is $0.016418 lower for this example. That is a small but visible difference, and you can calculate it before committing a larger job.

Cache reads are worth checking too. For 1,000,000 cache-read tokens, DeepSeek V4 Flash costs $0.005556 on the price list. MiMo V2.6 Pro costs $0.002778, while DeepSeek V4 Pro costs $0.0333. If your workload repeats context, cache-read pricing can materially change the total, so keeping it in the spreadsheet is sensible.

Claude Sonnet 5.5 provides another worked example. For 1,000,000 input tokens plus 200,000 output tokens, the CheapRelay cost is $0.8903 plus $0.89, or $1.7803. Its official list input price is $2.00 and output price is $10.00, so the official total for the same example is $2.00 plus $2.00, or $4.00. The difference is $2.2197 for the example.

04 / From an empty account to the first billed call

The setup path is short because the billing model is prepaid rather than tied to a monthly fee. Create an account on CheapRelay at https://cheapapirelay.com. Every new account receives $0.09 in sign-up credit, usable on any model. That credit is enough to make a small test call and see the flow work before topping up.

After signing in, add funds to the prepaid balance. Then create an API key and configure the appropriate client with the CheapRelay base address, https://api.cheapapirelay.com. For Claude, use Claude Code or Claude Desktop version 2.1.285 or newer and choose the desired Claude model id. For Gemini, use Gemini CLI, the native /v1beta route and a Gemini model id. For GPT through Codex, or for a compatible model such as DeepSeek V4 Flash, use the /v1 route and its exact model id.

Send a short request. The first bill is calculated from the real tokens used, not from a package estimate. A 402 response means the balance is at or below zero, which gives you a clear signal to add funds. Each key is limited to 120 calls per minute, and each account can have 8 calls running at once. Those limits make the initial test predictable and keep traffic within the account boundaries.

You can inspect the balance before and after the test to confirm the arithmetic. If the call used 1,000 input tokens and 500 output tokens with Qwen3.7 Plus, for example, the token calculation is 1,000 divided by 1,000,000 multiplied by $0.1208, plus 500 divided by 1,000,000 multiplied by $0.3625. That equals $0.0001208 plus $0.00018125, or $0.00030205. Tiny test calls are cheap because the price list is per real token.

05 / Keep the balance flexible and the rewards separate

Prepaid does not mean disposable. The balance does not expire, and refunds are available for unused balance within 12 days of a top-up. The refund is the unused amount minus what was spent and the payment fee, so keeping track of the top-up date and current balance makes the process straightforward.

CheapRelay also has a two-level referral arrangement. A directly invited user generates a 20% reward based on what that user pays, while the second level contributes 6%. These rewards enter a withdrawable rewards balance rather than the prepaid spending balance. The smallest withdrawal is ¥20, and withdrawals use Alipay. Keeping the two balances separate makes it easier to see money available for API calls and money available for withdrawal.

The overall calculation is straightforward: published token prices, a prepaid balance with no monthly fee, and client routes that match each model family. Claude Code gives Claude access, Gemini CLI provides the native Gemini path, and Codex handles GPT through the compatible /v1 route. For a concrete benchmark, GPT-6.1 Sol output is $0.7375 per 1,000,000 tokens, 93% under its $10.00 official list output price. That is the number to carry into your next budget.