Claude Opus 5 pricing: $8.28 per 1M output tokens

01 / What Claude Opus 5 costs per million tokens on CheapRelay

Here is the line I would put on the bill for Claude Opus 5, taken straight from the CheapRelay price list. Per 1,000,000 tokens, input costs $1.66, output costs $8.28, and a cache read costs $0.1222. The model id to enter is claude-opus-5. That is the whole price: three lines, no monthly fee, and you pay for the tokens your code actually sends and receives.

The official list price for the same three lines is $5.00 input, $25.00 output, and $0.50 cache read per 1,000,000 tokens, converted at 7.2 yuan per US dollar. The arithmetic is simple. Input on CheapRelay is $1.66 against $5.00, so $1.66 divided by $5.00 is 33.2% of list. Output is $8.28 against $25.00, and $8.28 divided by $25.00 is 33.12% of list. Cache read is $0.1222 against $0.50, which is 24.44% of list. Put another way, the output line sits 67% under the official list price, the input line sits 67% under, and the cache read line sits 76% under.

One compatibility note matters more than any discount banner. Claude models on CheapRelay run only in the official Claude Code and Claude Desktop clients, version 2.1.285 or newer. They do not run in Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK, cURL, or any OpenAI-compatible setup. If your workflow already lives in Claude Code, you change the endpoint and the model id, and the pricing above applies.

02 / A worked example week with Claude Opus 5

Everything below is an example, not a promise about your usage. Suppose one developer runs Claude Opus 5 through Claude Code for five working days. Each day the job sends 120,000 input tokens, receives 25,000 output tokens, and reads 300,000 tokens from cache.

• ■ Input per day: 120,000 divided by 1,000,000 is 0.12 of a million. 0.12 multiplied by $1.66 is $0.1992.
• ■ Output per day: 25,000 divided by 1,000,000 is 0.025. 0.025 multiplied by $8.28 is $0.207.
• ■ Cache read per day: 300,000 divided by 1,000,000 is 0.30. 0.30 multiplied by $0.1222 is $0.03666.
• ■ Total per day: $0.1992 plus $0.207 plus $0.03666 is $0.44286.
• ■ Total for the example week: $0.44286 multiplied by 5 is $2.2143.

That is the number I care about: about $2.21 for the example week, or roughly $0.44 per working day. Now split the same week into requests. If the day contains 20 calls, each call carries about 6,000 input tokens, 1,250 output tokens, and 15,000 cache-read tokens. One call then costs $0.00996 of input, $0.01035 of output, and $0.001833 of cache read, which is $0.022143 per call. One hundred calls across the example week come to $2.2143, matching the total above.

Two more example comparisons make the arithmetic easier to carry into a budget. If the same example week used 600,000 input tokens and no output, the input line would be 0.60 multiplied by $1.66, or $0.996. If it used 125,000 output tokens and no input, the output line would be 0.125 multiplied by $8.28, or $1.035. Cache-heavy work is the quiet saver: 1,500,000 cache reads for the example week would cost 1.5 multiplied by $0.1222, which is $0.1833.

03 / How that example week compares with the official list price

Run the same week at the official list rates and the arithmetic gets more satisfying. Input: 600,000 tokens is 0.60 of a million, and 0.60 multiplied by $5.00 is $3.00. Output: 125,000 tokens is 0.125 of a million, and 0.125 multiplied by $25.00 is $3.125. Cache read: 1,500,000 tokens is 1.5 of a million, and 1.5 multiplied by $0.50 is $0.75. The official list total for the example week is $3.00 plus $3.125 plus $0.75, which is $6.875.

CheapRelay charges $2.2143 for that same example week. The gap is $6.875 minus $2.2143, which is $4.6607. Measured against list, $4.6607 divided by $6.875 is 67.8%. So the saving is not a rounded slogan; it falls out of three multiplications and one subtraction.

That also tells you where the leverage sits. Output is the largest line in the example week, so trimming output tokens or choosing a lower output rate moves the bill more than trimming input. Cache reads are already a small line at $0.1222 per million, so a cache-heavy project is mostly paying for its output. CheapRelay does not hide this structure: the price list shows the same three columns for every Claude and GPT model, so you can run the arithmetic before you commit to a model.

04 / Choosing a lower-priced line when Opus 5 is not required

Opus 5 is the right pick when quality matters more than pennies, but the price list gives you real alternatives. This comparison is also an example, using the same 125,000 output tokens from the example week.

• ■ Claude Opus 5 output: 0.125 multiplied by $8.28 is $1.035.
• ■ Claude Opus 5.5 output: 0.125 multiplied by $6.24 is $0.78. The difference against Opus 5 is $1.035 minus $0.78, which is $0.255 for the example week.
• ■ Claude Sonnet 5 output: 0.125 multiplied by $4.88 is $0.61. The difference against Opus 5 is $1.035 minus $0.61, which is $0.425.
• ■ Claude Sonnet 5.5 output: 0.125 multiplied by $4.45 is $0.55625.
• ■ Kimi K3 output: 0.125 multiplied by $4.67 is $0.58375.
• ■ Gemini 3.1 Pro output: 0.125 multiplied by $6.49 is $0.81125.
• ■ DeepSeek V4 Pro output: 0.125 multiplied by $3.03 is $0.37875.
• ■ GPT-6 Astra output: 0.125 multiplied by $2.82 is $0.3525, and its official list output is $50.00 per million, so it sits 94% under list.

The input side behaves similarly in the example week at 600,000 input tokens: Opus 5 costs 0.60 multiplied by $1.66, or $0.996. Claude Opus 5.5 costs 0.60 multiplied by $1.32, or $0.792. Claude Sonnet 5 costs 0.60 multiplied by $0.9764, or $0.58584. GPT-6 Astra costs 0.60 multiplied by $0.5639, or $0.33834, against an official list input of $10.00 per million.

Cache reads reward you for reusing context. For the example week at 1,500,000 cache-read tokens, Opus 5 costs 1.5 multiplied by $0.1222, or $0.1833. Opus 5.5 costs 1.5 multiplied by $0.0597, or $0.08955. Claude Sonnet 5 costs 1.5 multiplied by $0.0861, or $0.12915. GPT-6 Astra costs 1.5 multiplied by $0.0569, or $0.08535. Pick the model from the workload, then check the line that dominates your own traffic.

05 / How CheapRelay billing works in plain terms

CheapRelay is prepaid and pay-as-you-go. There is no monthly fee, and your balance does not expire, so an unused dollar is not a deadline you have to beat. Every new account receives a $0.09 sign-up credit that works on any model, which is enough to fire one small test call and see the response arrive before you top up.

The refund rule is also specific. If you top up and then change your mind, you can ask for a refund of the unused balance within 12 days of the top-up. The refund subtracts what you already spent and the payment fee, so the number you get back is the unused part, not the full transaction.

Referrals pay into a separate withdrawable rewards balance. A directly invited user pays you 20% of what they pay, and the second level pays 6%. The smallest withdrawal is ¥20, and withdrawals go through Alipay. Because the reward sits in a withdrawable balance rather than being folded into your API credit, you can see the two figures separately.

06 / Setup, clients, and the limits you will actually hit

The API base for CheapRelay is https://api.cheapapirelay.com, and the site is https://cheapapirelay.com. Claude models go through the official Claude Code or Claude Desktop client at version 2.1.285 or newer. Gemini models use the native protocol under /v1beta through Gemini CLI, never an OpenAI-compatible client. GPT models, including the Codex route, plus DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and the other listed models work with any OpenAI-compatible client under /v1, such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL.

Two operating limits are worth knowing. 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, so top up and try again rather than rewriting your client.

For Claude Opus 5 specifically, the arithmetic in this article reduces to three numbers you can check on the price list: $1.66 per million input tokens, $8.28 per million output tokens, and $0.1222 per million cache-read tokens. In the worked example week, that is $2.2143 instead of $6.875 at official list rates, with no monthly fee attached.