$0.1236 per 1M input tokens: MiniMax M3 on a relay service

01 / $0.1236 per 1M input tokens puts MiniMax M3 on a practical budget

MiniMax M3 is listed on the CheapRelay price list at $0.1236 per 1 million input tokens, $0.4917 per 1 million output tokens, and $0.025 per 1 million cache-read tokens. That is enough information to calculate a bill before committing to a day of development work.

The model id you enter is MiniMax-M3. MiniMax M3 belongs to the group of models that work with an OpenAI-compatible client under /v1. That means you can use tools such as Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, or cURL. Point the client at the CheapRelay API base shown on the site, select MiniMax-M3, and send the request through /v1.

CheapRelay uses a prepaid balance and charges for real tokens. There is no monthly fee, and the balance does not expire. A new account also receives $0.09 in sign-up credit, usable on any model. For a developer checking the arithmetic, this is a straightforward arrangement: add funds, use tokens, and subtract the actual usage from the balance.

02 / The before-and-after bill: $0.23430 becomes $0.22444

Here is an example workload for MiniMax M3: 1 million input tokens, 200,000 output tokens, and 100,000 cache-read tokens.

First, consider the before bill as if every input-side token were charged at the regular input rate. The arithmetic is:

• ■ 1,000,000 input tokens × $0.1236 = $0.12360
• ■ 200,000 output tokens × $0.4917 = $0.09834
• ■ 100,000 cache-read tokens × $0.1236 = $0.01236
• ■ Before total: $0.12360 + $0.09834 + $0.01236 = $0.23430

Now apply the listed cache-read rate to those 100,000 cached tokens:

• ■ 100,000 cache-read tokens × $0.025 = $0.00250
• ■ After total: $0.12360 + $0.09834 + $0.00250 = $0.22444
• ■ Example saving: $0.23430 − $0.22444 = $0.00986

That example saves $0.00986, or about 4.2% of the before bill. It is not a dramatic headline number, but it is arithmetic you can reproduce from the price list. The $0.09 sign-up credit is also worth placing beside the example: $0.09 ÷ $0.22444 is about 40.1% of that after total. The credit can cover a meaningful part of the first example session, subject to the remaining balance.

Cache reads matter here because repeated context can be read at $0.025 per 1 million tokens instead of $0.1236. The saving becomes larger as more of the workload consists of cached context. You still need a prompt-caching workflow that produces cache reads; the price list explains the rate, while the actual token counts come from your requests.

03 / Set up MiniMax M3 without changing your client habits

MiniMax M3 uses the OpenAI-compatible route under /v1, so the setup can follow the client you already use. Enter MiniMax-M3 as the model id and use the CheapRelay API base shown on the site. Keep the endpoint under /v1 rather than switching protocols.

It is worth separating MiniMax M3 from the other model families on the price list. Claude models are restricted to the official Claude Code and Claude Desktop clients, version 2.1.285 or newer; they do not use the OpenAI-compatible route. Gemini models use their native protocol under /v1beta through Gemini CLI, and the OpenAI-compatible route does not apply to them. GPT, DeepSeek, Kimi, GLM, Qwen, Grok, MiniMax, and the other compatible models use /v1. Choosing the correct route avoids a configuration problem disguised as a model problem.

For MiniMax M3 specifically, the client list is straightforward: Cursor, Cherry Studio, Cline, Roo Code, OpenCode, the OpenAI SDK, and cURL are all compatible choices. The model id remains MiniMax-M3 in each case. If you use an SDK, keep the base URL and route under /v1; if you use cURL, send the same model id through the same compatible endpoint.

CheapRelay also publishes operational limits so you can size your first test. Each key is limited to 120 calls per minute, and each account can run 8 calls at once. If a request returns HTTP 402, the account balance is at or below zero. That gives you a concrete troubleshooting signal rather than an ambiguous failure to diagnose.

04 / What MiniMax M3 costs against the other models on the price list

MiniMax M3 has no official list price in the published table, so the useful comparison is with the other models shown there. Its input price is $0.1236 per 1 million tokens. Qwen3.7 Plus is listed at $0.1208, a difference of $0.0028 per 1 million input tokens. MiMo V2.6 Pro is listed at $0.1736, which is $0.05 more per 1 million input tokens.

Output is where MiniMax M3 costs more than MiMo V2.6 Pro. MiniMax M3 charges $0.4917 per 1 million output tokens, while MiMo V2.6 Pro charges $0.3472. The difference is $0.1445 per 1 million output tokens. For an example 1 million output tokens, that is $0.4917 versus $0.3472. If your workload is dominated by input and cached context, MiniMax M3’s input and cache rates may fit your calculation better; if output volume dominates, the difference deserves a place in the budget.

CheapRelay’s price-first approach also includes published gaps where official list prices exist. For example, GPT-6.1 Sol output is listed at $0.7375 per 1 million tokens against an official list price of $10.00. The arithmetic is $10.00 − $0.7375 = $9.2625, which is a 93% discount. GPT-5.6 Sol output is 91% under its $20.00 official list price, and GPT-6 Astra output is 94% under its $50.00 official list price. These are the gaps shown for models that have official list prices, not estimates for MiniMax M3.

05 / Prepaid funds, referral rewards, and refunds on the same account

The billing model stays simple after setup. You pay for real tokens from a prepaid balance, there is no monthly fee, and unused funds do not expire. If you top up and change your mind, an unused balance can be refunded within 12 days of the top-up. The refund subtracts what has already been spent and the payment fee.

CheapRelay also includes a referral program. A directly invited user contributes 20% of what they pay, while the second level contributes 6%. Those percentages land in a withdrawable rewards balance. The smallest withdrawal is ¥20, and withdrawals are handled through Alipay. This makes referrals useful for spreading the tool without turning the pricing explanation into a subscription pitch.

For a first-day plan, keep the accounting visible. Record the MiniMax M3 input, output, and cache-read counts, multiply each by the published rates, and compare the result with the prepaid balance. The example above ends at $0.22444 for 1 million input tokens, 200,000 output tokens, and 100,000 cache-read tokens. If the workload grows, the same three prices let you rebuild the estimate without guessing.

06 / A first-day MiniMax M3 budget with no monthly commitment

MiniMax M3 through CheapRelay is easy to evaluate because the important numbers fit on one short bill. Input is $0.1236 per 1 million tokens, output is $0.4917, and cache reads are $0.025. The compatible client options and the MiniMax-M3 model id remove guesswork, while prepaid billing lets you control the budget directly.

The example saving from using the published cache-read rate is $0.00986 on a $0.23430 workload. The $0.09 sign-up credit covers about 40.1% of the resulting $0.22444 example bill. Those are modest figures, but they are checkable figures, and checkable figures are the point: know the rates, count the tokens, watch the balance, and let the bill explain the value.