$0.5639 per 1M input tokens: GPT-6 Astra setup on CheapRelay

01 / What does GPT-6 Astra cost on CheapRelay?

Q: What is the price list for GPT-6 Astra today?

A: On CheapRelay, GPT-6 Astra is $0.5639 per 1,000,000 input tokens, $2.82 per 1,000,000 output tokens and $0.0569 per 1,000,000 cache-read tokens. The model id to enter is gpt-6-astra.

Q: How far below the official list price is that?

A: The official list price is $10 for input, $50 for output and $1 for cache reads. Here is the arithmetic:

• ■ Input: $10 - $0.5639 = $9.4361 saved per 1M tokens, so $9.4361 divided by $10 = 94.361% below list.
• ■ Output: $50 - $2.82 = $47.18 saved per 1M tokens, so $47.18 divided by $50 = 94.36% below list.
• ■ Cache reads: $1 - $0.0569 = $0.9431 saved per 1M tokens, so $0.9431 divided by $1 = 94.31% below list.

Q: What does that look like in yuan?

A: The published official prices are shown at 7.2 yuan per US dollar. For this example, 1M input tokens cost $10 on the official list, which is 10 multiplied by 7.2 = ¥72. Through CheapRelay, the same 1M input tokens cost $0.5639, which is 0.5639 multiplied by 7.2 = ¥4.06008. For 1M output tokens, the official list example is $50 multiplied by 7.2 = ¥360, while CheapRelay charges $2.82 multiplied by 7.2 = ¥20.304. For 1M cache-read tokens, the official list example is $1 multiplied by 7.2 = ¥7.20, while the CheapRelay example is $0.0569 multiplied by 7.2 = ¥0.40968.

Q: Is GPT-6 Astra the only GPT model on the price list?

A: No. GPT-6.1 Sol is $0.1486 input, $0.7375 output and $0.0347 for cache reads per 1M tokens. GPT-5.6 Sol is $0.3056 input, $1.71 output and $0.0264 for cache reads. GPT-6 Astra costs more than both on each listed token type, but its published official list price is also higher: $10 input and $50 output. The point is not that every model has the same price. The point is that you can see the number before you spend and compare it directly with list.

02 / How do I complete the GPT-6 Astra setup on CheapRelay?

Q: What do I need before the first request?

A: Create a CheapRelay account, add funds to the prepaid balance and copy the API key. The API base is https://api.cheapapirelay.com. Every new account includes $0.09 in sign-up credit, and that credit can be used on any model, including GPT-6 Astra.

Q: What base URL and model should I enter?

A: Use https://api.cheapapirelay.com as the base URL and gpt-6-astra as the model. GPT-6 Astra belongs to the GPT family, which uses the OpenAI-compatible route under /v1. That means you can use an OpenAI-compatible client under /v1, including Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK or cURL. GPT models can also be used through Codex.

Q: Which clients should I avoid for GPT-6 Astra?

A: The client rule for GPT models is straightforward: use an OpenAI-compatible client under /v1. The restrictions on Claude and Gemini clients do not apply to GPT-6 Astra. Claude models work only with the official Claude Code and Claude Desktop clients, version 2.1.285 or newer, and Gemini models use their native protocol under /v1beta through Gemini CLI. Keeping those paths separate avoids the usual setup mistake: copying a Gemini or Claude configuration into a GPT request and wondering why the protocol does not match.

Q: What should I test first?

A: Send one short request, then look at the balance. This example assumes 1,000 input tokens and 500 output tokens. At the published rates, input costs 1,000 divided by 1,000,000 multiplied by $0.5639 = $0.0005639. Output costs 500 divided by 1,000,000 multiplied by $2.82 = $0.00141. Add them and the example total is $0.0019739. That is the arithmetic behind the first bill, not a subscription price.

Q: Is there a monthly fee?

A: No monthly fee. CheapRelay uses a prepaid balance and charges per real token. There is no monthly charge to maintain, and the balance does not expire.

03 / How do I calculate a realistic GPT-6 Astra bill?

Q: How much would one million input and output tokens cost?

A: Here is a labelled example. If a workload uses 1M input tokens and 1M output tokens, the input line is $0.5639 and the output line is $2.82. The example total is $0.5639 + $2.82 = $3.3839 per million of each token type. At the published conversion of 7.2 yuan per dollar, that is $3.3839 multiplied by 7.2 = ¥24.36408.

Q: How much does cache reading change the total?

A: Here is another example. Suppose the same 1M-token workload includes 1M cache-read tokens. The input line is $0.5639, output is $2.82 and cache reads are $0.0569. The example total is $0.5639 + $2.82 + $0.0569 = $3.4408. In yuan at 7.2 per dollar, the example total is $3.4408 multiplied by 7.2 = ¥24.77376. Cache reads add $0.0569 to that example, so the published cache-read rate is worth checking before a long session.

Q: Is that lower than the official list for the same example?

A: Yes. For the 1M-input plus 1M-output example, official list pricing would be $10 + $50 = $60, or ¥432 at 7.2 yuan per dollar. The CheapRelay example is $3.3839, or ¥24.36408. The difference is $60 - $3.3839 = $56.6161 per example workload.

Q: How does GPT-6 Astra compare with other models on the same price list?

A: Here is a labelled example using 1M input and 1M output tokens. GPT-6.1 Sol costs $0.1486 + $0.7375 = $0.8861. GPT-6 Astra costs $0.5639 + $2.82 = $3.3839. In that example, Astra is $3.3839 divided by $0.8861 = 3.82 times the GPT-6.1 Sol example cost. If your workload does not need Astra’s tier, the arithmetic lets you test a lower-priced model first without changing your billing model.

Q: What about other OpenAI-compatible choices?

A: The price list also includes DeepSeek V4 Flash at $0.2389 input and $0.9556 output, Qwen3.7 Plus at $0.1208 input and $0.3625 output, and MiMo V2.6 Pro at $0.1736 input and $0.3472 output. For the same labelled 1M-plus-1M example, those totals are $1.1945, $0.4833 and $0.5208 respectively. GPT-6 Astra’s example total is $3.3839. Compare the lines your workload actually produces, then pick the model that fits the bill.

04 / How do prepaid balances, refunds and referral rewards work?

Q: How does prepaid billing reduce risk?

A: You top up the balance and CheapRelay charges per real token. There is no monthly fee, and unused balance does not expire. That makes the exposure visible: the balance shows what you have, while the price list shows what each token type costs.

Q: Can I get my top-up back?

A: Yes, within 12 days of the top-up. The refund is the unused balance, minus what was spent and minus the payment fee. For example, if a $20 top-up has $3.3839 spent from the 1M-input-plus-1M-output example above, the pre-fee unused amount is $20 - $3.3839 = $16.6161. The actual refund also subtracts the payment fee.

Q: What do referrals pay?

A: A directly invited user generates a 20% referral reward based on what that user pays. A second-level referral generates 6%. These rewards go into a separate withdrawable rewards balance. The smallest withdrawal is ¥20, and withdrawals are made through Alipay.

Q: Can the referral reward be used as API credit?

A: The rewards balance is withdrawable, with a ¥20 minimum via Alipay. Plan for it as rewards money rather than assuming it is the same as prepaid API credit. The API balance and the rewards balance are described separately.

Q: What does the $0.09 sign-up credit buy in practice?

A: Here is an example. At GPT-6 Astra’s output rate, $0.09 divided by $2.82 multiplied by 1,000,000 equals about 31,915 output tokens if the whole credit were used only for output. If it were used only for input, $0.09 divided by $0.5639 multiplied by 1,000,000 equals about 159,609 input tokens. Real requests mix token types, so the actual amount will differ. The credit is usable on any model, which also gives you room to test alternatives.

05 / What limits and errors should I expect on day one?

Q: Is there a request limit?

A: Each key is limited to 120 calls per minute, and each account can run 8 calls at once. Those are the published limits, so they are enough to plan a normal first session without guessing at hidden quotas.

Q: What does HTTP 402 mean?

A: HTTP 402 means the balance is at or below zero. Add funds to the prepaid balance, then retry. It is a balance signal, not a model or client error.

Q: Why did my bill differ from a rough estimate?

A: Check the three lines: input, output and cache reads. Here is a labelled example with 2M input tokens, 200,000 output tokens and 500,000 cache-read tokens. Input is 2 multiplied by $0.5639 = $1.1278. Output is 0.2 multiplied by $2.82 = $0.564. Cache reads are 0.5 multiplied by $0.0569 = $0.02845. The example total is $1.1278 + $0.564 + $0.02845 = $1.72025.

Q: What is the practical first-day routine?

A: Create the account, claim the $0.09 credit, top up only what you are comfortable testing, connect an OpenAI-compatible client under /v1, and enter gpt-6-astra. Send a short request, check the balance, then run one larger labelled example. Compare that bill with the published GPT-6 Astra rates: $0.5639 input, $2.82 output and $0.0569 cache reads per 1M tokens, against official list prices of $10, $50 and $1. CheapRelay makes the setup direct and the arithmetic visible, so you can decide what the model is worth before the balance gets used up.