MiniMax M3: relay service output costs $0.4917 per million tokens

01 / MiniMax M3 pricing starts with three visible numbers

If you are evaluating a relay service for MiniMax M3, the first useful question is not which dashboard looks busiest. It is what one million tokens will cost, line by line. On the CheapRelay price list today, MiniMax M3 is listed at $0.1236 per 1,000,000 input tokens, $0.4917 per 1,000,000 output tokens, and $0.025 per 1,000,000 cache-read tokens. Those are the three figures that determine your bill.

That separation matters because input and output are not interchangeable. A request that sends a short prompt and receives a long explanation has a different cost profile from a request that repeatedly reads a large cached context. CheapRelay charges for real tokens, uses a prepaid balance, and has no monthly fee. The balance does not expire. New accounts receive $0.09 in sign-up credit, usable on any model, so you can make a small test before committing to a larger top-up.

MiniMax M3 is available through the OpenAI-compatible route under /v1. That means you can use an OpenAI-compatible client such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL. The API base is https://api.cheapapirelay.com, and the model id to enter is MiniMax-M3. There is one operational detail to keep in mind: each key is limited to 120 calls per minute, and each account can run 8 calls at once. If you receive HTTP 402, the account balance is at or below zero.

02 / Scenario one: a small assistant request beside two alternatives

The following example uses three models from the current price list and assumes 10,000 input tokens plus 2,000 output tokens per request. It is an example, not a benchmark. The arithmetic uses the published per-million prices and divides each figure by 1,000,000.

• ■ MiniMax M3: 10,000 input tokens cost 10,000 ÷ 1,000,000 × $0.1236 = $0.001236. The 2,000 output tokens cost 2,000 ÷ 1,000,000 × $0.4917 = $0.0009834. The example total is $0.0022194.
• ■ Qwen3.7 Plus: 10,000 input tokens cost 10,000 ÷ 1,000,000 × $0.1208 = $0.001208. The 2,000 output tokens cost 2,000 ÷ 1,000,000 × $0.3625 = $0.000725. The example total is $0.001933.
• ■ MiMo V2.6 Pro: 10,000 input tokens cost 10,000 ÷ 1,000,000 × $0.1736 = $0.001736. The 2,000 output tokens cost 2,000 ÷ 1,000,000 × $0.3472 = $0.0006944. The example total is $0.0024304.

For this small-assistant pattern, Qwen3.7 Plus produces the lower example total: $0.001933 compared with MiniMax M3 at $0.0022194. The difference is $0.0002864 per request. MiMo V2.6 Pro comes to $0.0024304, which is $0.000211 more than MiniMax M3 in this example. The point is not that one model is always the right choice. The point is that CheapRelay lets you compare the actual price rows before you build a workflow around them.

These three models are all available through the OpenAI-compatible /v1 route. For MiniMax M3, use the model id MiniMax-M3. For Qwen3.7 Plus, use qwen3.7-plus. For MiMo V2.6 Pro, use MiMo-V2.6-Pro. The same client setup can serve several models while the cost calculation stays transparent.

03 / Scenario two: a longer working session, priced line by line

Here is a second example with 50,000 input tokens and 10,000 output tokens per request. That is a more representative shape for a developer asking a model to inspect several files, reason through a task, and return a patch or explanation.

• ■ MiniMax M3: input is 50,000 ÷ 1,000,000 × $0.1236 = $0.00618. Output is 10,000 ÷ 1,000,000 × $0.4917 = $0.004917. The example total is $0.011097.
• ■ Qwen3.7 Plus: input is 50,000 ÷ 1,000,000 × $0.1208 = $0.00604. Output is 10,000 ÷ 1,000,000 × $0.3625 = $0.003625. The example total is $0.009665.
• ■ MiMo V2.6 Pro: input is 50,000 ÷ 1,000,000 × $0.1736 = $0.00868. Output is 10,000 ÷ 1,000,000 × $0.3472 = $0.003472. The example total is $0.012152.

Qwen3.7 Plus is $0.001432 lower than MiniMax M3 in this example. MiniMax M3 is $0.001055 lower than MiMo V2.6 Pro. Multiply the example total by 100 identical requests and you get $1.1097 for MiniMax M3, $0.9665 for Qwen3.7 Plus, and $1.2152 for MiMo V2.6 Pro. That is the sort of calculation that makes a prepaid balance easy to plan: no fixed monthly charge, just usage multiplied by the price row you can see.

MiniMax M3’s output rate is $0.4917, compared with $0.3625 for Qwen3.7 Plus and $0.3472 for MiMo V2.6 Pro. Its input rate is $0.1236, compared with $0.1208 and $0.1736 respectively. The total tells you which row fits this particular traffic pattern. If your workload is output-heavy, compare the output lines; if it sends large prompts, compare the input lines.

04 / Scenario three: when cache reads become part of the bill

The third example adds 1,000,000 cache-read tokens to 1,000,000 input tokens and 100,000 output tokens. This is an example of a cache-heavy session. It is not a claim about how any particular client will behave; it simply shows how the three published MiniMax M3 lines combine.

• ■ MiniMax M3: input is $0.1236, output is 100,000 ÷ 1,000,000 × $0.4917 = $0.04917, and cache read is $0.025. The example total is $0.19777.
• ■ Qwen3.7 Plus: input is $0.1208, output is 100,000 ÷ 1,000,000 × $0.3625 = $0.03625, and cache read is $0.0361. The example total is $0.19315.
• ■ MiMo V2.6 Pro: input is $0.1736, output is 100,000 ÷ 1,000,000 × $0.3472 = $0.03472, and cache read is $0.002778. The example total is $0.211098.

Here, MiniMax M3’s cache-read line is $0.025 per million, while MiMo V2.6 Pro’s is $0.002778. The difference for one million cache-read tokens is $0.022222. MiMo’s lower cache-read figure offsets its higher input figure in this example: $0.211098 versus MiniMax M3 at $0.19777. Qwen3.7 Plus totals $0.19315, which is $0.00462 below MiniMax M3 for this mix.

That is why a single “price per million” headline is not enough. CheapRelay publishes input, output, and cache-read prices separately, so you can build the estimate that matches your code. If your application reads a large stable context repeatedly, include the cache line. If it mostly sends short prompts, do not inflate the estimate with a cache assumption.

05 / How to use MiniMax M3 without paying for idle features

The setup is deliberately narrow. Create a CheapRelay account, add funds to the prepaid balance, point an OpenAI-compatible client at https://api.cheapapirelay.com under /v1, and select the model id MiniMax-M3. The price list is the source for the three unit prices: $0.1236 input, $0.4917 output, and $0.025 cache read per 1,000,000 tokens.

There is no monthly fee. Usage is charged per real token, and unused balance does not expire. The $0.09 sign-up credit can be used on any model, which is enough to run a basic request while you verify your client configuration and your own arithmetic. If you top up and then decide not to use the account, unused balance can be refunded within 12 days of the top-up, minus what was spent and the payment fee.

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, with a ¥20 minimum withdrawal through Alipay. This is separate from your prepaid usage balance, so referral earnings do not confuse the amount available for API calls.

06 / Choosing the row that fits your actual traffic

MiniMax M3’s price row is easy to summarize: $0.1236 for input, $0.4917 for output, and $0.025 for cache reads, each per 1,000,000 tokens. In the three examples above, it produced totals of $0.0022194, $0.011097, and $0.19777. Qwen3.7 Plus and MiMo V2.6 Pro provide useful side-by-side figures, and the winner changes with the token mix. That is exactly why the price list is published in separate lines.

For developers who want a straightforward relay service with no monthly fee, the practical method is simple: estimate input, output, and cache reads separately, multiply each by the displayed rate, and compare the totals. CheapRelay gives you the unit prices and the access route; your workload determines the final bill. Sign-up credit, prepaid billing, and the published MiniMax M3 rates make the first test inexpensive to run and the next projection easy to audit.