MiniMax M3 and a Telegram Bot on Fluxio's Pay-as-You-Go API

§1 Day One: A Telegram Bot That Starts Rewriting Itself

Imagine you spend the first week of building a Telegram bot that does something slightly unusual: it begins sending its answer, then edits that answer as the response is still arriving. The bot receives a question in Telegram, calls Fluxio with MiniMax M3, and forwards each incoming token to the chat while it continues to improve the wording. The interesting part is not the concept. It is how naturally the flow works when the API is designed for streaming from the start.

Fluxio is a streaming-first, pay-per-token AI API relay. Replies arrive token by token over server-sent events, and every stream reports its real usage. For a Telegram bot, that means the first visible fragment can appear while the rest of the reply is still being generated. The bot can show a draft, revise a sentence, replace a phrase, and settle on a final version without waiting for a complete buffer.

On day one, the setup is deliberately small. The bot uses an OpenAI-format client pointed at the Fluxio API base, api.fluxioapi.com, with the /v1 path added as usual. MiniMax M3 is listed as available from any OpenAI-format client, including OpenAI SDKs and cURL, so it fits this kind of integration directly. You are not forced into a special editor or a bespoke client just to see a reply arrive in pieces.

§2 Day Two: Following the Token Bill Instead of Guessing

By the second day, the diary entries are about numbers. Fluxio bills each request on the tokens reported by the model, with four token types priced separately: input, output, cache write, and cache read. There is no flat per-request fee. That detail changes how you think about a bot that streams several revisions, because the amount of output you actually generate is the amount that gets counted.

MiniMax M3 is priced at input $0.1514, output $0.6028, and cache read $0.0306 per 1M tokens. The list of models available through Fluxio is broad enough that you can compare options while you build. Gemini 2.5 Flash, for example, is input $0.2097, output $1.68, cache read $0.0208, and cache write $0.2097 per 1M tokens. MiMo V2.6 Pro is input $0.2111, output $0.4208, and cache read $0.002778 per 1M tokens. DeepSeek V4 Flash is input $0.2875, output $1.15, and cache read $0.006944 per 1M tokens.

For example, if you send 100 messages a day, the useful question is not what a request costs in the abstract but how many tokens those 100 messages consume. Once the bot is live, you can read that directly in the Fluxio usage log, which records each request, its tokens, and its cost. That log turns a diary entry into a practical habit: check the day's output volume, then decide whether the bot should stream a shorter draft, keep a cached context, or trim its self-edits.

Pay-as-you-go also means there is no monthly fee on the balance plan. Top-ups are $1.39, $6.94, $13.89, $27.78, $69.44, and $138.89 through Stripe Checkout. If you are still experimenting in week one, a small top-up is enough to keep the diary honest. A new account also receives $0.12 in sign-up credit, which is a modest start for the first few test conversations.

§3 Day Four: The Connection Drops and the Reply Keeps Going

Day four contains the incident every streaming project eventually meets. A Telegram user closes the app, a proxy hiccups, or the bot's outbound request to Fluxio is interrupted mid-answer. With many APIs, that moment raises an awkward question about what you owe for a reply the user never saw in full.

Fluxio's behavior here is straightforward. If the client disconnects mid-stream, Fluxio still reads the reply to its end and bills the tokens actually used, nothing more. It never adds a penalty. Tokens that were already generated are not refunded, which is the honest counterpart to the rule: you are charged for the work that happened, and you are not charged for a hidden extra. For your Telegram bot, that means a dropped connection does not create a mysterious line item. You can look at the usage log and see the tokens and cost for that request.

This is also why the streaming design suits an editing bot. The reply is produced progressively, so a partial display is not a broken display. Your bot can post the first sentence, then send an edited version beneath it. The user watches the answer take shape, and the Fluxio stream keeps reporting real usage as it goes.

The account limits are sensible for a first-week project: 120 requests a minute per key, and 8 streams open at once per account. A single Telegram bot with one key can comfortably run several conversations at the same time, and you can raise the number of open streams as your usage log shows where the traffic actually sits.

§4 Day Five: Choosing a Model for Each Kind of Edit

By day five, the diary starts to include a routing table. MiniMax M3 is a natural default for a bot that writes quickly and revises often, because its output price is $0.6028 per 1M tokens. If a message needs a longer, more deliberate rewrite, you can look at another model in the Fluxio catalog and compare prices side by side.

Claude Fable 5 is input $3.50, output $17.50, cache read $0.2694, and cache write $4.31 per 1M tokens. Claude Sonnet 5.5 is input $1.19, output $5.97, cache read $0.1042, and cache write $1.49 per 1M tokens. Claude Opus 5.5 is input $1.51, output $7.13, cache read $0.0681, and cache write $1.78 per 1M tokens. GPT-6.1 Sol is input $0.1944, output $0.9722, and cache read $0.0389 per 1M tokens, while GPT-5.6 Sol is input $0.3861, output $2.16, and cache read $0.0306 per 1M tokens.

There are also monthly plans if your bot's traffic settles into a rhythm. The Claude monthly plan is $62.36 for a $1100 allowance of official value over 30 days, at most $350 a week, or $165.14 for a $2900 allowance over 30 days, at most $850 a week. The GPT monthly plan is $59.58 for a $2600 allowance of official value over 30 days, at most $800 a week, or $183.19 for a $8000 allowance over 30 days, at most $2400 a week. Each request draws its official-dollar value multiplied by 2.8, and the Claude plans cover Claude models only with their own API key, while the GPT plans cover GPT models only with their own API key.

One client note matters for this diary. Claude models run only in the official Claude Code and Claude Desktop apps, version 2.1.285 or newer, and they do not run in Cursor, Cherry Studio, Cline, OpenCode, any SDK, or cURL. Gemini models work only through their native protocol in Gemini CLI, and they are not OpenAI-compatible. MiniMax M3, GPT, DeepSeek, Kimi, GLM, Grok, Qwen, Hunyuan, and MiMo models work from any OpenAI-format client, including Cursor, Cherry Studio, Cline, OpenCode, OpenAI SDKs, and cURL, with Codex running GPT models. For a Telegram bot built on an OpenAI-format client, that distinction decides the model list on day one.

§5 Day Seven: A Small Diary, a Steady Stream

By the end of the first week, the diary has become a working system. The Telegram bot calls Fluxio, receives MiniMax M3 tokens as server-sent events, posts the opening lines, and edits its reply while the stream is still live. The usage log shows each request, its tokens, and its cost. The pay-as-you-go balance reflects what was actually consumed, with no flat per-request fee added on top.

Fluxio supports six interface languages: English, Chinese, Spanish, Hindi, Portuguese, and French. If something needs attention, messages sent from the site contact page or the dashboard appear in the dashboard, where replies are shown; there is no support email to hunt for. If you top up and then find the balance unused, you can ask through the message page for a refund within 4 days of that top-up, minus what was consumed and payment fees.

Referrals are part of the same balance-based picture: 17% of what a directly invited friend pays, 5% for the next level, with rebates able to move into the balance or be withdrawn from $12.50 by Alipay. The first-week diary ends with a clear picture of what streaming costs and what it delivers. The bot edits its reply because Fluxio streams it that way, and the bill follows the tokens that were really used.