MiniMax M3 API relay: pay-as-you-go with Relay365

The shift log opens with one key

The Relay365 duty desk opens a shift log for every project that comes through the door. This one belongs to a small side project: a helper that drafts short release notes and keeps their tone consistent. The first entry is not a subscription and not a contract. It is a key, created in the dashboard, tagged for the project, and bound to the MiniMax M3 lane. Relay365 is an independent third-party relay, run like a round-the-clock duty desk, and the desk keeps every lane on one table, so a small build does not need a procurement exercise before it can send a single request.

Everything the first afternoon needs sits in the same place. The API base is one endpoint, quick-wiring snippets sit next to the key list, and the key itself is tied to the lane you chose. A new account also arrives with $0.25 in signup credit, which is enough to make a first test call before reaching for a card. When the desk logs that first request, the entry is already visible in the usage log: the time, the model, the input, the output, the cache recall, the charge, and the result. Nothing about the call is a black box, and nothing about the account requires a monthly plan to keep open.

MiniMax M3 was the natural pick for this project because the desk publishes its token rates in the open, and the release-note helper needs a model that can write a paragraph and keep a voice, not a fleet of features. So the log entry for day one is short: one key, one lane, one test call, one visible charge.

MiniMax M3 rates on the desk board

Here is what the board says about MiniMax M3. Rates are per 1 million tokens in approximate USD, while the account itself is kept in CNY. Input is $0.1528, output is $0.6111, and cache read is $0.0306; no cache write rate is listed for this lane. The point of a published board is that a developer can do the arithmetic before writing the code, rather than discovering the shape of the bill afterwards.

On the desk, all four token types - input, output, cache write and cache read - are priced separately. A call is therefore charged for what it actually did, not for one blended guess, and a project that leans on cached context can see that difference in the log instead of feeling it as a mystery. This separation is one of the quiet reasons a relay can be easy to budget: the same four columns apply across the table, so the shape of one call tells you how the next call will be measured.

The example is deliberately plain. For example, if you send 1,000,000 input tokens and 100,000 output tokens a day on MiniMax M3, the input side costs $0.1528, and the output side costs $0.6111 for one tenth of a million, which is $0.06111. Together that is $0.21391 a day, and about $6.4173 over 30 days. This is an example only, not a quote for your project, and the number moves with the token counts your code actually sends. The useful part is the method: pick your daily token counts, multiply by the two published rates, and you have a working budget line before the first invoice.

Reading the usage log line by line

Every request can be traced in the usage log, which shows the latest 200 requests with the time, model, input, output, cache recall, charge and result of each call. For a project like this one, that log becomes the shift log itself. When a draft comes out clipped or a tone drifts, the desk can point to the exact call, the exact token counts and the exact charge, and the developer can see whether the fix belongs in the prompt or in the model choice.

One honest detail is worth stating plainly: the usage log does not show which key made the call. The desk still records the ledger of wallet, bounty and pass accounts, and the interface offers six interface languages, so the numbers can be read in the language the team prefers. In-site messages to the desk are available once you log in, which means a question about a specific line can be asked in the same place the line already lives.

The desk also states the operating limits plainly. An HTTP 429 appears when a single key makes more than 120 calls in 60 seconds, or when one account keeps more than 8 requests in flight. Those numbers are part of the shift description, not a footnote: a small project that suddenly fans out should know where the ceiling sits before it hits it. Keeping every lane on one table means the limits, the log and the charges are all visible from the same screen.

From the first bill to the first top-up

Relay365 bills from a prepaid wallet, with no monthly fee. Every call is settled on the token counts the upstream reports, so the charge in the log and the charge on the wallet are the same event seen from two angles. There is no fixed desk rent sitting on the account between months, no minimum activity, and no monthly line to justify. You top up what you intend to use, and the log tells you how much of it has been used.

Two billing details make that model easier to trust in practice. A call that fails upstream, with a 502 or 503, is not billed. And a disconnected stream is billed for what was actually used, so a connection that drops mid-response does not turn into a full-price request. Both rules are written down, which means the desk can answer a billing question with the rule instead of an opinion.

Refunds are handled the same way. Within 14 days of a top-up, unused balance can be refunded minus what was consumed and the payment channel's fees; bounty and promotional credit are not refundable. The referral bounty is 18% at level 1 and 6% at level 2 of what referred members pay, and it can be withdrawn from $11.11 or moved into the wallet. For a project that grows by word of mouth, that is a way to turn a colleague's usage into working credit rather than into a side conversation.

When the project needs a second lane

Then the release-note helper grows a second feature, and the log starts showing more than one model. GPT-6.1 Sol joins the table at $0.1986 input and $0.9944 output per 1 million tokens, with cache read at $0.0389; its output rate is about 10% of the official list price of $10. Claude Fable 5 sits at $3.79 input and $18.94 output, with cache write $4.66 and cache read $0.2917, and its output rate is about 38% of the official list price of $50. Claude models work only from the official Claude Code and Claude Desktop clients, with Claude Code 2.1.285 or newer. Cursor, Cherry Studio, Cline, OpenCode, SDKs and cURL pair with the other supported lanes instead, and MiniMax M3 and GPT-6.1 Sol are among them.

Gemini has its own lane and its own native protocol, reached through Gemini CLI: :generateContent for single replies and :streamGenerateContent?alt=sse for streams. It cannot go through /v1/chat/completions, so it is wired the way Gemini expects rather than forced into another shape. Different lanes, different rules, one table - that is the desk's job, and it saves a developer from relearning the plumbing every time the project changes direction.

For heavier Claude or GPT traffic, there are optional duty passes. The Claude pass at 14 days costs $50.97 and draws up to $700 of quota at the duty-pass rate x2.5, capped at $380 of quota per 7 days. The GPT pass at 30 days costs $81.53 and draws up to $3000 of quota at the duty-pass rate x2.5, capped at $900 of quota per 7 days. A pass comes with a dedicated key that never touches the wallet, and other models always use the wallet. The pricing page shows the full table of lanes, rates and passes, so you can pick the lane that matches the project instead of guessing.

The shift log for this small project now reads like any other: a first key, a first call, a first bill, and a desk that keeps the numbers visible. Create a Relay365 account, bind a key to the lane your project needs, and start the log with a test request.