Qwen3.8 Max pricing on CheapRelay: $0.5722 input, $1.72 output

01 / The myth: Qwen3.8 Max pricing is just one vague number

Model pricing becomes confusing when a provider shows one headline rate and leaves the actual bill to the imagination. The useful question is simpler: what do input tokens, output tokens, and cache-read tokens cost on the CheapRelay price list?

For Qwen3.8 Max, the model id to enter is qwen3.8-max. The current prices are $0.5722 per 1,000,000 input tokens, $1.72 per 1,000,000 output tokens, and $0.0681 per 1,000,000 cache-read tokens. That is the line-by-line schedule. There is no need to guess how a mixed request will be billed.

Here is an example. Suppose one request uses 1,000,000 input tokens, 1,000,000 output tokens, and 1,000,000 cache-read tokens. The arithmetic is straightforward:

• ■ Input: 1,000,000 ÷ 1,000,000 × $0.5722 = $0.5722.
• ■ Output: 1,000,000 ÷ 1,000,000 × $1.72 = $1.72.
• ■ Cache read: 1,000,000 ÷ 1,000,000 × $0.0681 = $0.0681.
• ■ Example total: $0.5722 + $1.72 + $0.0681 = $2.3603.

That is an example, not a package or a preset allowance. CheapRelay bills for real tokens against a prepaid balance. You are not forced to buy a fixed amount of usage, and there is no monthly fee.

02 / The myth: a lower rate on one token type settles the comparison

Another common myth is that the smallest number on a price list automatically makes a model the right choice. Token types do different work, so the meaningful comparison is the same workload priced on both models.

Compare Qwen3.8 Max with Qwen3.6 Plus Uncensored on the price list. Qwen3.8 Max costs $0.5722 per 1,000,000 input tokens and $1.72 per 1,000,000 output tokens. Qwen3.6 Plus Uncensored costs $0.3431 per 1,000,000 input tokens and $1.83 per 1,000,000 output tokens.

The input difference is $0.5722 − $0.3431 = $0.2291 per 1,000,000 input tokens. The output difference runs the other way: $1.83 − $1.72 = $0.11 per 1,000,000 output tokens. In this example, Qwen3.8 Max costs more on input and less on output. Which side matters depends on your traffic, which is why the separate rates are worth keeping visible.

Now compare Qwen3.8 Max with MiMo V2.6 Pro. MiMo V2.6 Pro is $0.1736 for input, $0.3472 for output, and $0.002778 for cache reads per 1,000,000 tokens. Qwen3.8 Max is $0.5722 for input, $1.72 for output, and $0.0681 for cache reads. For one million tokens of each type, the differences are:

• ■ Input: $0.5722 − $0.1736 = $0.3986.
• ■ Output: $1.72 − $0.3472 = $1.3728.
• ■ Cache read: $0.0681 − $0.002778 = $0.065322.

These are examples from today’s price list. They show why “one price” is not enough context: a heavy-output workload and a cache-heavy workload can produce very different totals.

03 / The myth: an official list-price comparison applies to every model

Official list prices are published on the price list for Claude and GPT models. No official list price is shown for Qwen3.8 Max, so it would be misleading to attach one to it. The honest approach is to compare Qwen3.8 Max with the other listed models and to use official list prices where they actually exist.

For example, GPT-6.1 Sol has an official output price of $10.00 per 1,000,000 tokens. On CheapRelay, its output price is $0.7375. The example saving is $10.00 − $0.7375 = $9.2625 per 1,000,000 output tokens, or 93% under the official list price.

GPT-5.6 Sol offers another published comparison. Its official output price is $30.00 per 1,000,000 tokens, while its CheapRelay output price is $1.71. The arithmetic is $30.00 − $1.71 = $28.29, which is 94% under list.

Claude Sonnet 5.5 has an official output price of $15.00 per 1,000,000 tokens. Its output price here is $4.45, so the example difference is $15.00 − $4.45 = $10.55, or 70% under the official list price. These comparisons belong to those models. They do not create an official list price for Qwen3.8 Max.

For Qwen3.8 Max itself, the defensible statement is precise: the published rates are $0.5722 input, $1.72 output, and $0.0681 cache read per 1,000,000 tokens. Price first, then the real comparison against the figures that are actually published.

04 / The myth: cache reads and client setup make the bill complicated

Cache reads have their own rate, and using one can make the arithmetic clearer rather than murkier. In the Qwen3.8 Max row, a cache read costs $0.0681 per 1,000,000 tokens, compared with $0.5722 for input.

Here is an example request containing 500,000 input tokens, 250,000 output tokens, and 2,000,000 cache-read tokens:

• ■ Input: 0.5 × $0.5722 = $0.2861.
• ■ Output: 0.25 × $1.72 = $0.43.
• ■ Cache read: 2 × $0.0681 = $0.1362.
• ■ Example total: $0.2861 + $0.43 + $0.1362 = $0.8523.

Setup is also less involved than many people expect. Qwen models use the OpenAI-compatible route under /v1. The API base is https://api.cheapapirelay.com. You can use an OpenAI-compatible client such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL.

One warning matters: do not apply that instruction to every model. Claude models work only with the official Claude Code and Claude Desktop clients, version 2.1.285 or newer, and never through Cursor, Cherry Studio, Cline, Roo Code, OpenCode, an SDK, cURL, or an OpenAI-compatible connection. Gemini models use the native protocol under /v1beta through Gemini CLI, never the OpenAI-compatible route. For Qwen3.8 Max, the OpenAI-compatible path is the relevant one.

05 / The myth: prepaid billing means extra charges or wasted balances

The final myth is that a prepaid system must include hidden overhead or expire without warning. CheapRelay uses a prepaid balance with pay-per-real-token billing, no monthly fee, and no expiration on the balance. The amount you add is the amount available for token usage.

Every new account receives $0.09 in sign-up credit, usable on any model. That is small enough to treat as a practical test: send a request, watch the token charges, and see the balance move. It is credit for real usage, not a separate subscription.

CheapRelay also has a referral program. A directly invited user contributes 20% of what they pay, while the second level contributes 6%. Those amounts go into a withdrawable rewards balance. The smallest withdrawal is ¥20, paid through Alipay.

The refund rule is direct as well. Unused balance can be refunded within 12 days of the top-up, minus what was spent and the payment fee. There is no monthly fee to offset, no expiration clock on the balance, and no need to buy more tokens than your workload can use.

Finally, the operating limits are stated plainly: 120 calls per minute per key and 8 concurrent calls per account. An HTTP 402 response means the balance is at or below zero, so you know exactly when to top up.

Put the arithmetic together and the decision becomes manageable. Qwen3.8 Max costs $0.5722 per 1,000,000 input tokens, $1.72 per 1,000,000 output tokens, and $0.0681 per 1,000,000 cache-read tokens on the CheapRelay price list. Compare those three numbers with your own token mix, check the other models where a closer fit may cost less, and use the published list-price gaps where they exist. That is a budget you can calculate before you make the call.