DeepSeek V4 Pro: $1.01 per 1M input on CheapRelay API relay

01 / The bill starts with three published rates

Here is the part most price pages make you work for: CheapRelay shows the rate before you spend money, so you can calculate the bill instead of guessing at it. DeepSeek V4 Pro has three rates on the price list, all measured per 1,000,000 real tokens.

• ■ Input tokens cost $1.01 per 1,000,000 tokens.
• ■ Output tokens cost $3.03 per 1,000,000 tokens.
• ■ Cache-read tokens cost $0.0333 per 1,000,000 tokens.

That is the whole pricing formula. There is no monthly fee, no seat charge and no bundle to decode. CheapRelay uses a prepaid balance and charges per real token. The balance does not expire, which means money left on the account remains available for later work rather than disappearing at the end of a billing cycle.

The model id you enter is deepseek-v4-pro. DeepSeek models use the OpenAI-compatible route under /v1, so you can use clients such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK or cURL. The API base is https://api.cheapapirelay.com. In other words, the same price applies across the client path: you are not buying a special client plan or a separate subscription.

New accounts also receive $0.09 in sign-up credit, usable on any model. For a small first test, that credit can cover a request. For a working session, the arithmetic below shows a more useful bill.

02 / Before the request: the balance you bring

Imagine a before-and-after bill. Before the request, you top up $10. If the $0.09 new-account credit is applied, the available prepaid balance is $10.09. No monthly fee is attached to that balance, and it does not expire.

Now consider this example workload:

• ■ 1,000,000 input tokens
• ■ 500,000 output tokens
• ■ 1,000,000 cache-read tokens

Every line follows the same arithmetic: tokens used divided by 1,000,000, multiplied by the published rate.

Input: 1,000,000 ÷ 1,000,000 × $1.01 = $1.01.

Output: 500,000 ÷ 1,000,000 × $3.03 = $1.515.

Cache read: 1,000,000 ÷ 1,000,000 × $0.0333 = $0.0333.

After the request, the example bill is $1.01 + $1.515 + $0.0333 = $2.5583.

Starting from $10.09, the remaining balance is $10.09 − $2.5583 = $7.5317. That is the after-the-request bill in full: $2.5583 spent and $7.5317 left for the next request. Nothing needs a subscription renewal, and the leftover balance carries forward.

This example also makes the output premium visible. Output is $3.03 per million, while input is $1.01 per million. The arithmetic does not hide that difference; it shows exactly where the money goes.

03 / After the request: a smaller bill, line by line

Here is a second example with a lighter workload: 100,000 input tokens, 20,000 output tokens and 50,000 cache-read tokens.

Input: 100,000 ÷ 1,000,000 × $1.01 = $0.101.

Output: 20,000 ÷ 1,000,000 × $3.03 = $0.0606.

Cache read: 50,000 ÷ 1,000,000 × $0.0333 = $0.001665.

The example total is $0.101 + $0.0606 + $0.001665 = $0.163265, or $0.1633 when rounded to four decimal places. If the $0.09 sign-up credit covers this example, the credit exceeds the example charge by $0.09 − $0.163265, so this particular workload still requires a small amount of prepaid funds.

That is the useful property of per-token billing: the bill changes with the work. A short request produces a small charge; a long session produces a larger one. You can estimate either one from the three published rates before sending it.

Cache reads are worth watching because their rate is much lower than input. In the second example, 50,000 cache-read tokens cost $0.001665, while 50,000 input tokens would cost 50,000 ÷ 1,000,000 × $1.01 = $0.0505. The difference is $0.0505 − $0.001665 = $0.048835 for that example. The arithmetic is straightforward, so cache-heavy usage can be evaluated without hand-waving.

04 / How DeepSeek V4 Pro compares with V4 Flash

The price list includes DeepSeek V4 Flash, so the comparison can stay on the same model family. DeepSeek V4 Flash charges $0.2389 for input, $0.9556 for output and $0.005556 for cache read per 1,000,000 tokens. DeepSeek V4 Pro charges $1.01, $3.03 and $0.0333 respectively.

For each line, the example difference is straightforward:

• ■ Input: $1.01 − $0.2389 = $0.7711 per million tokens.
• ■ Output: $3.03 − $0.9556 = $2.0744 per million tokens.
• ■ Cache read: $0.0333 − $0.005556 = $0.027744 per million tokens.

Those figures describe the rate difference, not a verdict about which model fits a task. If a workload needs the capabilities of V4 Pro, the relevant question is what that workload costs. Using the same example mix as before, V4 Flash costs 1,000,000 ÷ 1,000,000 × $0.2389 = $0.2389 for input, plus 500,000 ÷ 1,000,000 × $0.9556 = $0.4778 for output, plus 1,000,000 ÷ 1,000,000 × $0.005556 = $0.005556 for cache reads.

The V4 Flash example total is $0.2389 + $0.4778 + $0.005556 = $0.722256. The V4 Pro example total is $2.5583. The difference is $2.5583 − $0.722256 = $1.836044 for this example workload.

CheapRelay also publishes official list prices for Claude and GPT models, so its price-first claim can be checked against a published benchmark. For example, GPT-6.1 Sol output is $0.7375 against an official list price of $10.00, which is 93% under list. Claude Sonnet 5 output is $4.45 against an official list price of $15.00, which is 70% under list. For DeepSeek V4 Pro, the price list does not show an official list price, so the honest comparison is the published V4 Flash rate above.

05 / Setting up DeepSeek without changing your workflow

DeepSeek V4 Pro is configured through the OpenAI-compatible path. Use the model id deepseek-v4-pro, point the client at the https://api.cheapapirelay.com base and use the /v1 path. A client such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK or cURL can use that route.

The account limits are also stated plainly: 120 calls per minute per key and 8 calls at once per account. If a request returns HTTP 402, the balance has reached or passed zero. That response gives you a concrete signal to top up instead of leaving you to infer what happened from a generic error.

Refunds follow a defined rule as well. Unused balance can be refunded within 12 days of the top-up, with the amount spent and the payment fee deducted. That makes the top-up decision easier to review: the refund window and the calculation are visible in advance.

CheapRelay also has a referral program. You receive 20% of what a directly invited user pays and 6% from a second-level referral. These payments go into a withdrawable rewards balance, with a minimum withdrawal of ¥20 through Alipay. That is separate from your prepaid API balance, so it does not change the token prices calculated above.

06 / What the final bill tells you

DeepSeek V4 Pro costs $1.01 per million input tokens, $3.03 per million output tokens and $0.0333 per million cache-read tokens on CheapRelay. The example workload of 1,000,000 input, 500,000 output and 1,000,000 cache-read tokens produced a $2.5583 bill. The smaller example produced $0.163265 before rounding.

The before-and-after is easy to check: $10.09 available, $2.5583 spent, $7.5317 remaining. There is no monthly fee, the prepaid balance does not expire, and every charge comes from real token counts multiplied by a rate you can see in advance.

For developers, that means the budget follows the code. You know the model id, the endpoint path, the client options, the rate limits and the cost of each token category. CheapRelay makes the arithmetic part of the purchase decision rather than something you discover on the invoice.