Claude Sonnet 5.5 API relay: check the $0.8903 rate before you pay

01 / The before-and-after bill for Claude Sonnet 5.5

The price is only useful when you put it beside the same bill elsewhere. That is why I would check Claude Sonnet 5.5 on CheapRelay against its official list price before adding funds. The price list shows $0.8903 per million input tokens, $4.45 per million output tokens, and $0.0778 per million cache-read tokens. The official list prices for the same three categories are $2.00, $10.00, and $0.10. Those are the figures that matter; a headline price without matching token categories is not a comparison.

CheapRelay also gives every new account a $0.09 sign-up credit that can be used on any model. Billing is prepaid and pay-as-you-go: you pay per real token, there is no monthly fee, and the balance does not expire. That makes the arithmetic straightforward. You can calculate a workload, compare the two bills, and decide how much to top up without buying a subscription or committing to a recurring charge.

02 / Before: what the official list-price bill looks like

Let us use one clearly labelled example: a workload containing 1,000,000 input tokens, 1,000,000 output tokens, and 1,000,000 cache-read tokens. At the official list prices, the input line is 1,000,000 divided by 1,000,000 multiplied by $2.00, so it costs $2.00. The output line is $10.00. The cache-read line is $0.10. The total is $2.00 + $10.00 + $0.10 = $12.10.

The same example expressed in yuan, using the stated conversion of 7.2 yuan per US dollar, is $2.00 × 7.2 = ¥14.40 for input, $10.00 × 7.2 = ¥72.00 for output, and $0.10 × 7.2 = ¥0.72 for cache reads. The combined bill is ¥87.12. This is the “before” side of the comparison: a bill built entirely from the published official list prices.

Now keep that example visible. The useful question is not whether a relay exists or whether a discount is advertised. It is whether the after bill is smaller for the same token counts, and whether the saving can be reproduced with plain arithmetic.

03 / After: the same workload on CheapRelay

On the CheapRelay price list, the same 1,000,000 input tokens cost $0.8903. The same 1,000,000 output tokens cost $4.45. The same 1,000,000 cache-read tokens cost $0.0778. The total is $0.8903 + $4.45 + $0.0778 = $5.4181.

Subtract that from the official-list example: $12.10 − $5.4181 = $6.6819. The after bill is therefore $6.6819 lower for this example workload. In yuan, the CheapRelay lines are $0.8903 × 7.2 = ¥6.41016, $4.45 × 7.2 = ¥32.04, and $0.0778 × 7.2 = ¥0.56016. The combined total is ¥39.01016, compared with ¥87.12 in the before example. The difference is ¥48.10984.

Look at the output line alone because output is where the gap is clearest. The official list price is $10.00 per million tokens; the CheapRelay price is $4.45. The saving is $10.00 − $4.45 = $5.55 per million output tokens, which is about 56% under the official list price. Input goes from $2.00 to $0.8903, a saving of $1.1097 per million tokens. Cache reads go from $0.10 to $0.0778, a saving of $0.0222 per million tokens.

04 / How the gap compares with other published prices

One number does not explain a price list by itself, so here is another labelled example. Claude Sonnet 5.5 lists output at $4.45 per million tokens. Claude Sonnet 5 lists output at $4.88. Since $4.45 is $0.43 below $4.88, Sonnet 5.5 has the lower published output price. For input, Sonnet 5.5 is $0.8903 while Sonnet 5 is $0.9764, a difference of $0.0861 per million tokens. For cache reads, the figures are $0.0778 and $0.0861, a difference of $0.0083.

The same arithmetic works across the other Claude entries. Claude Opus 5.5 lists input at $1.32 and output at $6.24, compared with Sonnet 5.5 at $0.8903 input and $4.45 output. The differences are $0.4297 for input and $1.79 for output. That comparison tells you where the price increases when moving up the named range; it does not require an assumed list price for any unlisted model.

The GPT rows provide another published comparison because their official list prices are also shown. GPT-6.1 Sol is $0.1486 input and $0.7375 output, against official list prices of $2.00 and $10.00. GPT-5.6 Sol is $0.3056 input and $1.71 output, against $4.00 and $20.00. GPT-6 Astra is $0.5639 input and $2.82 output, against $10.00 and $50.00. These figures make the gap visible without turning it into a vague promise.

05 / Using Claude Sonnet 5.5 with the correct client

Easy to use starts with putting the model in the right place. Claude models on CheapRelay work only with the official Claude Code client and the official Claude Desktop client, version 2.1.285 or newer. The model id to enter is claude-sonnet-5-5. Claude models are not connected through an OpenAI-compatible route, so using that route for Sonnet 5.5 is not the supported setup.

By contrast, the GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and other listed models use an OpenAI-compatible client under /v1. Examples include Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK and cURL. Gemini models use their native protocol under /v1beta through Gemini CLI. The API base for CheapRelay is https://api.cheapapirelay.com.

The service also publishes practical operating limits: 120 calls per minute per key and 8 concurrent calls per account. If you receive HTTP 402, the balance is at or below zero, so topping up is the next action rather than changing clients or rewriting the request.

06 / What happens to the money after you top up

Prepaid does not mean trapped. If you have unused balance within 12 days of a top-up, you can request a refund, minus the amount already spent and the payment fee. The balance does not expire, and there is no monthly fee, so unused funds can remain available rather than forcing another purchase.

CheapRelay also runs a two-level referral programme. A directly invited user’s payments generate 20% for you, while the second level generates 6%. Those amounts go into a withdrawable rewards balance, with the smallest withdrawal set at ¥20 through Alipay. Those are concrete percentages rather than vague rewards language.

Put the two bills together and the decision becomes measurable. In the example above, the official-list total was $12.10, while the CheapRelay total was $5.4181, a difference of $6.6819. The model id is claude-sonnet-5-5, the supported clients are official Claude Code and Claude Desktop at version 2.1.285 or newer, and the billing remains prepaid, per-token, and free of a monthly fee. That is the arithmetic worth checking before you pay.