DeepSeek V4 Pro pricing: $1.01 per 1M input tokens

01 / Start with the arithmetic: DeepSeek V4 Pro costs $1.01 per 1M input tokens

If you are comparing API prices for real work, the useful question is not whether one number feels impressive. It is what your bill looks like after the tokens are counted. CheapRelay publishes token prices in dollars per 1,000,000 tokens, so you can multiply the model rate by the number of tokens your request actually uses. The price list starts with smaller figures for some models and rises as the model class changes. There is no monthly fee: you fund a prepaid balance, and the balance does not expire.

Here is the first receipt line for the model in the title. DeepSeek V4 Pro has an input price of $1.01 per 1,000,000 tokens, an output price of $3.03 per 1,000,000 tokens, and a cache read price of $0.0333 per 1,000,000 tokens. The model id to enter is deepseek-v4-pro. That is enough to begin a comparison without guessing.

• ■ Step 1 — Buy DeepSeek V4 Pro input — $1.01 per 1M tokens.
• ■ Step 2 — Buy DeepSeek V4 Pro output — $3.03 per 1M tokens.
• ■ Step 3 — Buy DeepSeek V4 Pro cache reads — $0.0333 per 1M tokens.
• ■ Step 4 — Fund a new account with the sign-up credit — $0.09, usable on any model.

Example scenario: a request that uses 100,000 input tokens and 20,000 output tokens on DeepSeek V4 Pro costs $0.101 for input plus $0.0606 for output, or $0.1616 in total. The calculation is simply $1.01 multiplied by 0.1, plus $3.03 multiplied by 0.02. The cache read rate applies only when the same cached material is read again, so include it only when that feature is part of the request.

02 / Compare the price list with official Claude and GPT prices

The official list comparison is available for Claude and GPT models. At the published exchange rate of 7.2 yuan per US dollar, the list price for GPT-6.1 Sol input is $2.00 per 1M tokens, output is $10.00, and cache read is $0.10. CheapRelay lists GPT-6.1 Sol at $0.1486 input, $0.7375 output, and $0.0347 cache read. The output figure is 93% under the official list price: $10.00 minus $0.7375 equals $9.2625, and $9.2625 divided by $10.00 equals 92.625%, which rounds to the published 93% gap.

• ■ Step 5 — Compare GPT-6.1 Sol input — $0.1486 versus official $2.00 per 1M tokens.
• ■ Step 6 — Compare GPT-6.1 Sol output — $0.7375 versus official $10.00 per 1M tokens.
• ■ Step 7 — Compare GPT-6.1 Sol cache reads — $0.0347 versus official $0.10 per 1M tokens.
• ■ Step 8 — Compare GPT-5.6 Sol output — $1.71 versus official $20.00 per 1M tokens; the published gap is 91% under list.
• ■ Step 9 — Compare GPT-6 Astra output — $2.82 versus official $50.00 per 1M tokens; the published gap is 94% under list.

Claude shows the same kind of visible arithmetic. Claude Haiku 4.5 is $0.2917 input, $1.40 output, and $0.025 cache read per 1M tokens, compared with official list prices of $1.00, $5.00, and $0.10. Its output price is 72% under list: $5.00 minus $1.40 equals $3.60, and $3.60 divided by $5.00 equals 72%.

• ■ Step 10 — Compare Claude Haiku 4.5 input — $0.2917 versus official $1.00 per 1M tokens.
• ■ Step 11 — Compare Claude Haiku 4.5 output — $1.40 versus official $5.00 per 1M tokens.
• ■ Step 12 — Compare Claude Sonnet 5.5 output — $4.45 versus official $10.00 per 1M tokens; the published gap is 56% under list.
• ■ Step 13 — Compare Claude Sonnet 5 output — $4.88 versus official $10.00 per 1M tokens; the published gap is 51% under list.
• ■ Step 14 — Compare Claude Opus 5 output — $8.28 versus official $25.00 per 1M tokens; the published gap is 67% under list.

Example scenario: 50,000 output tokens on Claude Sonnet 5 cost $0.244 on CheapRelay, because $4.88 multiplied by 0.05 equals $0.244. At the official $10.00 output rate, the same 50,000 tokens cost $0.50. The difference for that example is $0.256.

03 / Compare other models by their actual token rates

For models without an official list price shown here, compare the CheapRelay price list with the other CheapRelay rows. That keeps the comparison grounded in figures that are published today. Qwen3.7 Plus, model id qwen3.7-plus, is $0.1208 input, $0.3625 output, and $0.0361 cache read. MiMo V2.6 Pro, model id MiMo-V2.6-Pro, is $0.1736 input, $0.3472 output, and $0.002778 cache read. MiniMax M3, model id MiniMax-M3, is $0.1236 input, $0.4917 output, and $0.025 cache read.

• ■ Step 15 — Compare Qwen3.7 Plus with MiMo V2.6 Pro output — $0.3625 versus $0.3472 per 1M tokens.
• ■ Step 16 — Compare Qwen3.7 Plus with MiniMax M3 input — $0.1208 versus $0.1236 per 1M tokens.
• ■ Step 17 — Compare DeepSeek V4 Flash with DeepSeek V4 Pro output — $0.9556 versus $3.03 per 1M tokens.
• ■ Step 18 — Compare Gemini 2.5 Flash with Gemini 3.1 Pro output — $1.42 versus $6.49 per 1M tokens.
• ■ Step 19 — Compare GLM-5.3 with Grok 4.7 output — $1.38 versus $1.52 per 1M tokens.
• ■ Step 20 — Compare Kimi K3 with Claude Sonnet 5.5 output — $4.67 versus $4.45 per 1M tokens.

Example scenario: a 200,000-token output run on DeepSeek V4 Flash costs $0.19112, because $0.9556 multiplied by 0.2 equals $0.19112. The same volume on DeepSeek V4 Pro would cost $0.606, because $3.03 multiplied by 0.2 equals $0.606. The comparison is between the two listed rates; no official list price is needed for either row.

04 / Pick the client that matches the model protocol

Low prices are useful only when the request reaches the model correctly. CheapRelay makes the client choice explicit. Claude models work only with the official Claude Code and Claude Desktop clients, version 2.1.285 or newer. Do not use Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK, cURL, or an OpenAI-compatible setup for Claude models.

• ■ Step 21 — Connect Claude models through Claude Code or Claude Desktop — $0.025 to $0.1222 cache read per 1M tokens, depending on the model.
• ■ Step 22 — Connect Gemini models through Gemini CLI using the native protocol under /v1beta — $0.0181 to $0.1125 cache read per 1M tokens.
• ■ Step 23 — Connect GPT models through an OpenAI-compatible client under /v1 — $0.0264 to $0.0569 cache read per 1M tokens.
• ■ Step 24 — Connect DeepSeek V4 Pro through an OpenAI-compatible client under /v1 — $0.0333 cache read per 1M tokens.
• ■ Step 25 — Connect GLM, Qwen, Grok, MiniMax, Kimi, and other listed models through an OpenAI-compatible client under /v1 — $0.002778 to $0.1264 cache read per 1M tokens.

The API base is https://api.cheapapirelay.com. The site is https://cheapapirelay.com. Gemini is native-protocol only through Gemini CLI, not an OpenAI-compatible client. GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax, and the other listed models can use an OpenAI-compatible client under /v1, including Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL.

05 / Check the balance, limits, referrals, and refund rule

CheapRelay bills against a prepaid balance, with payment based on real tokens and no monthly fee. The balance does not expire. Every new account receives $0.09 in sign-up credit, usable on any model. If the balance reaches zero or goes below it, the service returns HTTP 402, which tells you to top up before the next request. The rate limit is 120 calls per minute per key, and the account limit is 8 calls at once.

• ■ Step 26 — Add prepaid funds before the balance reaches $0 — $0.09 sign-up credit is available on a new account.
• ■ Step 27 — Track a key within its call allowance — 120 calls per minute per key.
• ■ Step 28 — Track account concurrency — 8 calls at once per account.
• ■ Step 29 — Invite a user directly — 20% of what that user pays goes to a withdrawable rewards balance.
• ■ Step 30 — Invite through the second level — 6% of what that user pays goes to the same rewards balance.
• ■ Step 31 — Request a refund for unused balance — request within 12 days of the top-up; the refund subtracts spending and the payment fee, with a minimum withdrawal of ¥20 via Alipay.

Example scenario: if a directly invited user pays $10, the referral reward is $2, because 20% of $10 equals $2. If a second-level invited user pays $10, the reward is $0.60, because 6% of $10 equals $0.60. These are referral examples, not token prices. The smallest withdrawal is ¥20 through Alipay.

The final check is simple: choose the model, read its input, output, and cache-read rates, select the matching client and protocol, and multiply by the tokens your request uses. CheapRelay gives you a prepaid balance, published per-token prices, and clear client boundaries. The arithmetic is visible before you spend, which is the point.