$10 Buys Claude Haiku 4.5 Tokens via a CheapRelay relay service

01 / The before bill: a subscription you keep paying for

Side projects rarely arrive on a tidy schedule. One week I am cleaning up a small reporting script, the next week I am rewriting a prompt for a weekend experiment, and then the project sits untouched for a month. That rhythm is a poor fit for a fixed recurring charge, because the bill keeps asking for money while the work is idle.

Here is the before bill, as an example. A developer signs up for a coding assistant subscription at $20 per month. In that month, the side project uses the assistant on four afternoons for roughly two hours each. The subscription does not care about those eight hours. It charges the full $20, and the $20 arrives whether the project shipped or not. Over a year, that is $240 for a tool that was genuinely useful for eight hours per month and then switched off.

CheapRelay takes the other approach. You add money to a prepaid balance, and you pay per real token. There is no monthly fee, and the balance does not expire. If a project stops for six weeks, the money simply waits. If it starts again, the balance is still there.

The prices are published on the price list, so the comparison does not depend on a discount that expires next week. Every new account also receives $0.09 in sign-up credit, usable on any model. That is small, and I will not pretend it is a month of coding. It is enough to run a first test and see the numbers on your own screen.

02 / What $10 buys on Claude Haiku 4.5

Claude Haiku 4.5 is a practical model for irregular side projects because the price list gives it a specific per-token rate. Here is an example, using $10 of prepaid balance and output tokens only, since output is usually the expensive side of a bill.

• ■ CheapRelay price for Claude Haiku 4.5 output: $1.40 per 1,000,000 tokens.
• ■ Official list price for the same output: $5.00 per 1,000,000 tokens.
• ■ $10 divided by $1.40 equals about 7.14 million output tokens.
• ■ $10 divided by $5.00 equals 2 million output tokens at official list price.
• ■ The gap is 7.14 million minus 2 million, or about 5.14 million extra output tokens, which matches the published 72% under list for output.

Now run the same arithmetic on input. The CheapRelay price for Claude Haiku 4.5 input is $0.2917 per 1,000,000 tokens, against an official list price of $1.00. So $10 buys $10 divided by $0.2917, which is about 34.28 million input tokens, compared with 10 million tokens at official list price. Cache reads are cheaper again: $0.025 per 1,000,000 tokens on CheapRelay, versus $0.10 at official list price.

That is the whole sales pitch in one line. Same model, same client, and a bill you can predict before you start. No monthly fee means the number on the bill is the number of tokens you actually sent.

For comparison within the price list, MiMo V2.6 Pro output is $0.3472 per 1,000,000 tokens and Qwen3.7 Plus output is $0.3625. MiniMax M3 output is $0.4917, which is $0.1445 above Qwen3.7 Plus. If your side project mostly writes text, picking the model by output rate is a straightforward calculation rather than a guess.

03 / What $10 buys on GPT-6.1 Sol and GPT-5.6 Sol

The GPT models show the same pattern with larger published gaps. Take GPT-6.1 Sol as an example. Output costs $0.7375 per 1,000,000 tokens on CheapRelay, while the official list price is $10.00. With $10 in prepaid balance, $10 divided by $0.7375 equals about 13.56 million output tokens. At official list price, $10 buys 1 million output tokens. The published figure for output is 93% under list.

GPT-5.6 Sol is priced at $1.71 per 1,000,000 output tokens on CheapRelay, against an official list price of $20.00. Here, $10 divided by $1.71 equals about 5.85 million output tokens, while $10 divided by $20.00 equals 0.5 million tokens at official list price. That is the published 91% under list for output. Input is $0.3056 per 1,000,000 tokens versus $4.00 at official list price.

GPT-6 Astra sits at $2.82 per 1,000,000 output tokens, with an official list price of $50.00 and a published 94% under list for output. Claude Sonnet 5.5 is $4.45 per 1,000,000 output tokens against $10.00 at official list price, a published 56% under list. Claude Opus 5.5 is $6.24 versus $20.00, published at 69% under list.

Put side by side, the arithmetic is easy to read. On output, GPT-6.1 Sol at $0.7375 is below GPT-5.6 Sol at $1.71, and Claude Haiku 4.5 at $1.40 is below Claude Sonnet 5.5 at $4.45. You can choose the model that fits the job and still see exactly what each million tokens costs.

04 / The after bill: one prepaid balance, the right client

The second half of the after bill is setup. CheapRelay is an API relay service at https://api.cheapapirelay.com, and the client you use depends on the model you selected. Get this wrong and you will spend an afternoon reading error messages.

• ■ For Claude models, including Claude Haiku 4.5, use only the official Claude Code or Claude Desktop clients, version 2.1.285 or newer. Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK and cURL are not the route for Claude models, and neither is an OpenAI-compatible setup.
• ■ For Gemini models, use the native protocol under /v1beta through Gemini CLI. Gemini models are not served through an OpenAI-compatible endpoint.
• ■ For GPT models, including Codex usage, and for DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax and the other models on the price list, use any OpenAI-compatible client under /v1, such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK or cURL.

When you enter a model, use the model id from the price list. Claude Haiku 4.5 is claude-haiku-4-5-20251001. GPT-6.1 Sol is gpt-6.1-sol. Gemini 2.5 Flash is gemini-2.5-flash. Qwen3.7 Plus is qwen3.7-plus. That removes guesswork.

There are two operating limits worth writing on the same sticky note as your API key: 120 calls per minute per key, and 8 calls at once per account. If you see HTTP 402, the balance is at or below zero. That is not a mysterious failure. It is a request to top up.

05 / How the prepaid balance behaves

A prepaid balance is easy to reason about because there is no renewal date and no monthly fee. You pay for the tokens you use, and the balance does not expire. For a side project that runs in bursts, that is the difference between paying for capacity and paying for activity.

The refund rule is also concrete. Unused balance can be refunded within 12 days of the top-up, minus what was spent and the payment fee. So the practical rule is simple: top up the amount you expect to use next, and check that first refund window if you are testing several models in one go.

Referrals add a small second-order option. You receive 20% of what a directly invited user pays, and 6% from the second level. Those rewards go into a separate withdrawable rewards balance. The smallest withdrawal is ¥20, paid through Alipay. It is not a subscription and it does not require a monthly fee, because there is no subscription in the first place.

06 / The after bill, line by line

Here is the complete after bill as an example. A developer tops up $10 on CheapRelay and uses $0.09 of the sign-up credit plus prepaid funds. The project runs Claude Haiku 4.5 through the official Claude Code client, version 2.1.285 or newer. The bill is per real token: $0.2917 per 1,000,000 input tokens, $1.40 per 1,000,000 output tokens, and $0.025 per 1,000,000 cache-read tokens.

Compare that with the before bill: a $20 recurring subscription charged every month, including the months where the project sleeps. With CheapRelay, a quiet month costs nothing beyond the balance already sitting there, and an active month costs only the tokens you actually used. The official list price for the same Claude Haiku 4.5 output would be $5.00 per 1,000,000 tokens, so the arithmetic behind the savings is visible on the price list rather than hidden in a contract.

That is the pitch I would make to another developer: show the rates, show the token counts, show the refund window, and let the arithmetic do the arguing. CheapRelay is a prepaid relay service with published per-token prices, no monthly fee, a balance that does not expire, and a client rule you can read once and write down. For irregular side projects, paying for what you use is the sensible default, and here the numbers make that case on their own.