$20 Buys Claude Opus 5.5 Output via a Cheaper API Relay

01 / The Myth: A Relay Must Charge Near the Official Claude List Price

Some developers assume that an API relay has to add a thick margin to a premium model, so a headline discount sounds fictional. That assumption is easy to test: take the published numbers and do the division.

On the CheapRelay price list today, Claude Opus 5.5 has an input price of $1.32 per 1,000,000 tokens, an output price of $6.24 per 1,000,000 tokens, and a cache read price of $0.0597 per 1,000,000 tokens. The official list price for the same categories is $4.00 input, $20.00 output, and $0.20 cache read.

Here is the arithmetic. $6.24 divided by $20.00 is 0.312, so the output price sits at roughly 31% of the official list price, which is the published 69% gap. $1.32 divided by $4.00 is 0.33, so input sits at one-third of list. $0.0597 divided by $0.20 is about 0.2985, so cache reads sit at roughly 30% of list. No estimation is required here. The price list already carries both sides of the comparison, and the gap is visible without a spreadsheet.

What does a fixed amount buy? Consider an example budget of $20 spent entirely on Claude Opus 5.5 output at the CheapRelay price. $20.00 divided by $6.24 is about 3.2 million output tokens. Spend the same $20 on input instead, and $20.00 divided by $1.32 is about 15.15 million input tokens. These are examples based on today's price list, not quotas or guarantees. They simply show what the money maps to.

02 / The Myth: Prepaid Access Means Paying for a Plan You May Not Use

The second myth is that a low per-token price only makes sense inside a long commitment. CheapRelay does not work that way. There is no monthly fee. You fund a prepaid balance, and you pay per real token. The balance does not expire, so unused funds are not tied to a renewal date.

New accounts also receive $0.09 in sign-up credit, usable on any model. That is a small number, and it is presented as a small number. It is enough to confirm that your client, your model id, and your key are wired correctly before you commit a larger balance.

The refund rule is similarly concrete. Unused balance can be refunded within 12 days of the top-up, minus what was spent and the payment fee. So the risk of testing a prepaid relay is bounded by the money you actually put in, not by a recurring charge.

Run the numbers on a simple example. Suppose you top up $20. At the Claude Opus 5.5 output price of $6.24 per 1,000,000 tokens, that $20 covers roughly 3.2 million output tokens as calculated above. If you only use part of it, the remainder stays on the prepaid balance, and if you decide within 12 days that you want it back, the unused portion can be refunded with the stated deductions. That is a different proposition from buying a subscription and discovering at the end of the cycle that you paid for capacity you never called.

03 / The Myth: A Cheap Relay Means Fighting the Client

The third myth assumes that a lower price comes with a complicated setup. On CheapRelay, the routing rules are published per model family, and they are short.

• ■ Claude models, including Claude Opus 5.5, work only with the official Claude Code and Claude Desktop clients, version 2.1.285 or newer. Not Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK, or cURL, and not an OpenAI-compatible setup.
• ■ Gemini models use the native protocol only, under /v1beta, through Gemini CLI.
• ■ GPT models, including GPT-6.1 Sol, GPT-5.6 Sol, and GPT-6 Astra, plus DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax, and the other models, use any OpenAI-compatible client under /v1, such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL.

For Claude Opus 5.5 specifically, the model id to enter is claude-opus-5-5. If you already work in Claude Code or Claude Desktop, the setup is the client you have, pointed at the CheapRelay base address https://api.cheapapirelay.com, with the model id filled in. If you are a Cursor user, you are not being asked to force Claude through an OpenAI-compatible route; you use the /v1 path with the OpenAI-compatible models instead.

The limits are also stated up front: 120 calls per minute per key, 8 calls at once per account. An HTTP 402 response means the balance is at or below zero, which is a signal to top up rather than a mystery failure. Nothing about this section requires invention. It is the routing table, read out loud.

04 / The Myth: A Discount on Output Does Not Survive a Real Request

The fourth myth is that advertised gaps evaporate once cache reads and input tokens are counted. Test the claim with one example request and one set of token counts.

Example: a request that consumes 100,000 input tokens, 10,000 output tokens, and 50,000 cache-read tokens on Claude Opus 5.5.

At the CheapRelay price list, input is $1.32 multiplied by 0.1, which is $0.132. Output is $6.24 multiplied by 0.01, which is $0.0624. Cache read is $0.0597 multiplied by 0.05, which is $0.002985. The example total is $0.132 plus $0.0624 plus $0.002985, or $0.197385.

At the official list price for the same token counts, input is $4.00 multiplied by 0.1, or $0.40. Output is $20.00 multiplied by 0.01, or $0.20. Cache read is $0.20 multiplied by 0.05, or $0.01. The official-list example total is $0.61.

The difference is $0.61 minus $0.197385, or $0.412615 in the example. The gap holds across all three token categories, because each one is priced below its official list figure rather than only the headline output rate.

It is also worth keeping the model rows separate. Claude Opus 5.5 is $1.32 input and $6.24 output. Claude Opus 5 is $1.66 input and $8.28 output, with an official list of $5.00 and $25.00, so its output is 67% under list. Claude Sonnet 5.5 is $0.8903 input and $4.45 output, with an official list of $2.00 and $10.00, so its output is 56% under list. Claude Sonnet 5 is $0.9764 input and $4.88 output, with an official list of $2.00 and $10.00, so its output is 51% under list. Claude Haiku 4.5 is $0.2917 input and $1.40 output, with an official list of $1.00 and $5.00, so its output is 72% under list. Every one of those comparisons uses only the two published figures for that same model.

05 / The Myth: Starting With CheapRelay Requires a Big Decision

The final myth is that a prepaid relay expects a large opening commitment before you learn anything. The mechanics say otherwise. You register, you receive $0.09 in credit usable on any model, and you test the route you actually intend to use.

For Claude Opus 5.5, the test is straightforward: open Claude Code or Claude Desktop at version 2.1.285 or newer, enter the model id claude-opus-5-5, and check the balance against the price list. Once you know your own token pattern, you can convert it into money with the same arithmetic used above.

One more practical note. The referral program pays 20% of what a directly invited user pays and 6% from the second level, credited to a withdrawable rewards balance with a ¥20 minimum withdrawal via Alipay. That is separate from API usage credit, and it does not alter the token prices.

Put the whole case in one line. Claude Opus 5.5 on CheapRelay costs $1.32 input, $6.24 output, and $0.0597 cache read per 1,000,000 tokens, against an official list of $4.00, $20.00, and $0.20. There is no monthly fee, the balance does not expire, and the price list does the selling for you. CheapRelay is registered and priced at cheapapirelay.com, and the remaining question is simply how many tokens your work actually consumes.