DeepSeek V4 Pro pricing puts one million output tokens at $3.03.
01 / CheapRelay account receipt: credit, prepaid balance and no monthly fee
Here is the CheapRelay bill before a developer sends the first request. The starting line is simple: prepaid balance, pay per real token, and no monthly fee. A new account receives $0.09 in sign-up credit, usable on any model. That credit is a practical way to run a small test before deciding which model belongs in production.
- ■ Sign-up credit: 1 account, $0.09, usable on any model.
- ■ Billing unit: 1 real token, charged from the prepaid balance.
- ■ Monthly charge: no monthly fee.
- ■ Balance expiry: no expiry date.
- ■ API base: https://api.cheapapirelay.com.
The receipt total for account setup is $0.09 in available credit. There is no separate subscription line, no recurring plan line, and no balance that quietly disappears on a monthly schedule. The balance does not expire, so a small remaining amount can stay available for the next request instead of turning into a wasted payment.
This structure also makes the refund question easier to understand. CheapRelay does not sell fixed time blocks of access. It records the balance you add, records the token usage charged against that balance, and keeps the remainder available. The important deadline is tied to the top-up: unused balance can be refunded within 12 days of the top-up, minus what was spent and the payment fee.
For a developer, that means two numbers should be kept near the account: the top-up amount and the usage total. Example only: if the amount added is T, token spending is U, and the payment fee is F, the refund calculation is T minus U minus F. The refund concerns the unused portion, not the full top-up after usage has already been deducted.
Keep the top-up receipt, check the usage figure, and review the refund within the 12-day window. That is the short version of the CheapRelay refund policy.
02 / Refund receipt: subtract usage and the payment fee from the top-up
The refund formula is short enough to fit on one line of a receipt. Start with the amount added in the top-up. Subtract what was spent. Subtract the payment fee. The result is the refundable unused balance, provided the request falls within 12 days of the top-up.
- ■ Line 1: top-up amount, quantity 1.
- ■ Line 2: less token spending already charged, quantity 1 usage total.
- ■ Line 3: less payment fee, quantity 1.
- ■ Total: refundable unused balance.
Example only: a request uses 1,000,000 input tokens and 1,000,000 output tokens on DeepSeek V4 Pro. The price list shows $1.01 per 1,000,000 input tokens and $3.03 per 1,000,000 output tokens. The receipt total is $1.01 plus $3.03, or $4.04. If the same request also reads 1,000,000 cached tokens, add $0.0333, producing $4.0733 in usage.
Now place that usage total into the refund line. Example only: if the top-up amount is T and the payment fee is F, the refund is T minus $4.04 minus F for the two-million-token request above. If cached tokens are included, the deduction becomes T minus $4.0733 minus F. The example shows the arithmetic; the actual refund follows the amount added, the recorded usage, and the fee on that top-up.
One distinction matters: the $0.09 new-account credit is sign-up credit, while the refund rule concerns a top-up and its unused balance. Also, spending does not prevent a refund by itself; spending reduces the amount. The 12-day top-up window determines whether the unused balance qualifies for a refund at all.
CheapRelay therefore gives developers a predictable calculation. No estimate is needed for a subscription that has already run out. The balance ledger tells you what was added, what was consumed, and what remains.
03 / Price-list receipt: token lines, published gaps and a visible total
Refunds are easier to reason about when the usage line itself is clear. CheapRelay publishes prices per 1,000,000 tokens, so a request can be turned into a receipt before it runs.
- ■ DeepSeek V4 Pro: 1,000,000 input tokens, $1.01; 1,000,000 output tokens, $3.03; 1,000,000 cache-read tokens, $0.0333.
- ■ GPT-6.1 Sol: 1,000,000 input tokens, $0.1486; 1,000,000 output tokens, $0.7375; 1,000,000 cache-read tokens, $0.0347.
- ■ Claude Sonnet 5: 1,000,000 input tokens, $0.9764; 1,000,000 output tokens, $4.88; 1,000,000 cache-read tokens, $0.0861.
- ■ Qwen3.7 Plus: 1,000,000 input tokens, $0.1208; 1,000,000 output tokens, $0.3625; 1,000,000 cache-read tokens, $0.0361.
For models with an official list price, the gap is also arithmetic. GPT-6.1 Sol output is $0.7375 against an official list output price of $10.00. The difference is $10.00 minus $0.7375, or $9.2625, which is 92.625% below the official list price, rounded to the published 93% gap. Its input line is $2.00 minus $0.1486, or $1.8514 below the $2.00 official list input price. Its cache-read line is $0.10 minus $0.0347, or $0.0653 below the $0.10 official list cache-read price.
Claude Sonnet 5 output is $4.88 against an official list output price of $10.00. The difference is $10.00 minus $4.88, or $5.12, which is 51.2% below list, matching the published 51% figure. Input is $2.00 minus $0.9764, or $1.0236 below the $2.00 official list input price. Cache reads are $0.20 minus $0.0861, or $0.1139 below the $0.20 official list cache-read price.
For models without an official list price on this page, compare their published lines with each other. Example only: a receipt with 1,000,000 input and 1,000,000 output tokens on DeepSeek V4 Pro totals $4.04. The same quantities on Gemini 3.1 Pro total $1.12 plus $6.49, or $7.61. The difference is $7.61 minus $4.04, or $3.57. This is a price-list comparison, not an estimate of another provider's bill.
04 / Client routing receipt: match the protocol to the model
The next line is configuration. CheapRelay provides a documented API base and a specific client route for each model family. Getting that line right avoids a failed request and keeps the refund arithmetic clean because failed setup work does not need to become part of the intended usage plan.
- ■ Claude models: use 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 setup.
- ■ Gemini models: use the native protocol under /v1beta through Gemini CLI. Do not use an OpenAI-compatible setup for Gemini.
- ■ GPT, Codex, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and the other listed models: use an OpenAI-compatible client under /v1. Examples include Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK and cURL.
Enter the model id exactly as published. For example, use deepseek-v4-pro for DeepSeek V4 Pro, qwen3.7-plus for Qwen3.7 Plus, gpt-6.1-sol for GPT-6.1 Sol, claude-sonnet-5 for Claude Sonnet 5, and gemini-3.1-pro for Gemini 3.1 Pro. The API base for these calls is https://api.cheapapirelay.com.
Example only: a developer choosing between GPT-6.1 Sol and Claude Sonnet 5 can write the model id first, then price the request from the price list. If the request uses 1,000,000 output tokens, the GPT-6.1 Sol line is $0.7375 and the Claude Sonnet 5 line is $4.88. The receipt difference is $4.88 minus $0.7375, or $4.1425. The published official list output prices are $10.00 for each of those models, so the savings gap can be checked directly.
That is the practical value of showing the numbers: model selection, client setup and expected token cost can sit on the same receipt before the request is sent.
05 / Reward and usage receipt: referrals, call limits and the final balance
CheapRelay also includes a rewards line for developers who bring colleagues. A directly invited user generates 20% of what that user pays. The second level generates 6%. These amounts go into a withdrawable rewards balance, where the smallest withdrawal is ¥20 through Alipay.
- ■ Direct referral rate: 20% of the directly invited user's payment.
- ■ Second-level referral rate: 6% of the second-level user's payment.
- ■ Rewards account: withdrawable balance.
- ■ Smallest withdrawal: ¥20, via Alipay.
Example only: using the published $0.09 sign-up credit as a payment amount, a direct referral produces 20% of $0.09, or $0.018. A second-level referral produces 6% of $0.09, or $0.0054. The example is scaled to the published credit amount; the account rule is the percentage applied to what the invited user actually pays.
There are two operating limits to include on the receipt. Each key permits 120 calls per minute, and each account permits 8 calls at once. An HTTP 402 response means the balance is at or below zero. That code is not a mysterious model failure; it is a balance signal telling you to add funds before the next request.
Now close the CheapRelay receipt. Account credit: $0.09. Billing: pay per real token. Monthly charge: no monthly fee. Balance expiry: none. Refund line: unused top-up balance within 12 days, less spending and the payment fee. Reward line: 20% direct, 6% second level, withdrawable from ¥20 through Alipay. Total: you can calculate the cost, the refund and the referral credit before committing the balance.
That arithmetic is the point. Start with the $0.09 credit, route the model through the correct client, watch the token lines, and keep the top-up receipt for the 12-day refund window. CheapRelay makes the bill visible enough to check.
More reading
| Date | Aisle | Written in | Title |
|---|---|---|---|
| 2026-11-03 | Cheap how-tos | EN | Kimi K3 API relay: 1M output tokens for $4.67 on CheapRelay Kimi K3 API relay walkthrough: referral rewards, prepaid billing and the first billed call on CheapRelay. |
| 2026-11-02 | Cheap how-tos | EN | $0.1208 per 1M tokens: CheapRelay Codex pay-as-you-go for students CheapRelay and Codex with pay-as-you-go pricing, token-level costs, signup credit, and clear client rules for students. |
| 2026-10-30 | Cheap how-tos | EN | $3 a Month Bought Our Team a Prompt Style Guide A beginner walkthrough for writing a team prompt style guide on a $3 monthly budget: templates, token math, tests and a weekly review. |
| 2026-10-30 | Cheap how-tos | EN | Trim Prompt Cost from $12.40 to $1.40 a Week A beginner walkthrough for cutting user data out of prompts, with worked token math that turns a $12.40 week into $1.40. |
| 2026-10-30 | Cheap how-tos | EN | A $10 Test Budget for Judging AI Output A beginner walkthrough for judging AI output quality with a spreadsheet, a rubric and a $10 budget. |
| 2026-10-30 | Cheap how-tos | EN | $3.80 a Month After Dropping the Free Tiers A beginner walkthrough for moving a hobby project from free AI tiers to pay-as-you-go on a five-dollar monthly budget. |