GPT-6.1 Sol relay service: check the $0.7375 output rate first

01 / Before you pay: what a relay service should show you

A relay service is useful only when the arithmetic is easy to inspect. For this example, I checked CheapRelay's price list, the model names, the client requirements, and the billing rules before adding any funds. The point is not to promise that every workload will cost the same amount. The point is to make the cost visible before the work starts.

CheapRelay bills by real token usage from a prepaid balance. There is no monthly fee, and the balance does not expire. Every new account also receives a $0.09 sign-up credit, usable on any model. The balance is practical for a light month because unused money is not tied to a recurring subscription. You top up, use what you need, and keep the remainder available for another request.

The price list is per 1,000,000 tokens. That unit matters when comparing a small experiment with a busy production day. I will use one worked week as an example, with clearly stated token volumes rather than vague claims about savings.

02 / Monday through Wednesday: three days of ordinary development

For this example, assume a developer uses one million input tokens, one million output tokens, and two million cache-read tokens each day for three days. The model is GPT-6.1 Sol, whose model id is gpt-6.1-sol. Its CheapRelay price list rates are $0.1486 for input, $0.7375 for output, and $0.0347 for cache read.

Here is the arithmetic for one day:

• ■ Input: 1,000,000 × $0.1486 = $0.1486.
• ■ Output: 1,000,000 × $0.7375 = $0.7375.
• ■ Cache read: 2,000,000 × $0.0347 = $0.0694.
• ■ Daily total: $0.1486 + $0.7375 + $0.0694 = $0.9555.

Three days in this example cost 3 × $0.9555 = $2.8665. That is a concrete figure for a concrete workload. If the subscription you barely open costs more than $2.87 for the same three days, CheapRelay is already cheaper for this example. If it costs less, the comparison may still change as usage grows, so the daily numbers are worth keeping.

For the same token volumes, the official list prices shown for GPT-6.1 Sol are $2.00 for input, $10.00 for output, and $0.10 for cache read. The official list total would be $2.00 + $10.00 + $0.20 = $12.20 per day. CheapRelay's $0.9555 is $11.2445 below that figure for the example day. The price list also shows GPT-6.1 Sol output at 93% under the official list price.

03 / Thursday: a Claude Code task with a different price

On Thursday in this example, the same developer switches to Claude Sonnet 5.5, whose model id is claude-sonnet-5-5. Claude models on CheapRelay use only the official Claude Code and Claude Desktop clients, version 2.1.285 or newer. They do not use Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK, or cURL, and they are not OpenAI-compatible endpoints.

Assume Thursday uses 500,000 input tokens, 500,000 output tokens, and 1,000,000 cache-read tokens. The price list rates are $0.8903 for input, $4.45 for output, and $0.0778 for cache read.

• ■ Input: 500,000 × $0.8903 = $0.44515.
• ■ Output: 500,000 × $4.45 = $2.225.
• ■ Cache read: 1,000,000 × $0.0778 = $0.0778.
• ■ Thursday total: $0.44515 + $2.225 + $0.0778 = $2.74795.

The official list prices for Claude Sonnet 5.5 are $2.00 for input, $10.00 for output, and $0.10 for cache read. At those rates, the same Thursday example would be $1.00 + $5.00 + $0.10 = $6.10. CheapRelay's $2.74795 is $3.35205 lower in this example. The published output rate is 56% under the official list price.

The client rule is part of the arithmetic too. If your daily tool is an OpenAI-compatible client such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL, it can be used with GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax, and the other models on the /v1 route. Gemini models are different: they use the native protocol under /v1beta through Gemini CLI, not an OpenAI-compatible client.

04 / Friday: comparing models without borrowing another list price

Not every model has an official list price on the price list, so I compare the listed CheapRelay rates directly. For another Friday example, assume 2,000,000 input tokens and 1,000,000 output tokens for a short script and a longer explanation.

With Qwen3.7 Plus, whose model id is qwen3.7-plus, the cost is 2 × $0.1208 + $0.3625 = $0.6041. With MiMo V2.6 Pro, whose model id is MiMo-V2.6-Pro, it is 2 × $0.1736 + $0.3472 = $0.6944. MiMo V2.6 Pro costs $0.0153 less on output for this example, while Qwen3.7 Plus costs $0.0528 less on input.

That is the useful part of a transparent price list: the choice can follow the shape of the task. If the script is mostly input, Qwen3.7 Plus is the lower figure. If the explanation is the larger portion, MiMo V2.6 Pro's $0.3472 output rate is lower than Qwen3.7 Plus's $0.3625. Both are simple prepaid calculations, with no subscription sitting unused in the background.

Other listed examples show the same approach. DeepSeek V4 Flash, model id deepseek-v4-flash, is $0.2389 for input, $0.9556 for output, and $0.005556 for cache read. Gemini 2.5 Flash, model id gemini-2.5-flash, is $0.1778 for input, $1.42 for output, and $0.0181 for cache read. For a one-million-input, one-million-output example, those are $1.1945 and $1.5978 respectively. The numbers make the trade-off visible without pretending that one model has a list price that the price list does not show.

05 / The example week, added up

For this worked example, the three GPT-6.1 Sol days cost $2.8665, and the Claude Sonnet 5.5 day costs $2.74795. The total is $5.61445 for the four days. If Friday uses the Qwen3.7 Plus example above, the week costs $5.61445 + $0.6041 = $6.21855.

Now compare the first four days with the official list prices shown for those two models. GPT-6.1 Sol would be 3 × $12.20 = $36.60, and Claude Sonnet 5.5 would be $6.10, making $42.70 for the same example workloads. CheapRelay's $5.61445 is $37.08555 lower for those four days. That is arithmetic on the listed rates, not a promise about every possible project.

There are a few practical details worth checking before paying. The API base is https://api.cheapapirelay.com, and the site is https://cheapapirelay.com. Each key is limited to 120 calls per minute, and each account to 8 calls at once. HTTP 402 means the balance is at or below zero. Unused balance can be refunded within 12 days of the top-up, minus spending and the payment fee.

CheapRelay also has a referral program: 20% of what a directly invited user pays, and 6% from the second level, goes into a withdrawable rewards balance. The smallest withdrawal is ¥20, via Alipay. For a developer who wants a small, measurable start, the combination of prepaid balance, per-token billing, published rates, and a $0.09 sign-up credit makes the first test easy to calculate: choose the model, count the tokens, and check the total.