When people talk about AI, the same names always come up: ChatGPT, Claude, Gemini. American heavyweights, all three. Yet one of the most interesting developments is happening elsewhere — and a model from there is actually working on this very blog.
We're talking about MiniMax, an AI provider from China. Its newest model, MiniMax M3, has only been on the market since early June — and here, it handles the editorial review of every new article before it goes live.
A Fraction of the Price
The real shocker is the price.
Through the API (the direct machine access, no app wrapped around it), MiniMax M3 currently costs around $0.30 per million input tokens and $1.20 per million output tokens — an introductory discount, by the way, that probably won't last forever. For comparison: Claude's flagship Opus 4.8 runs at $5 and $25 respectively. That's roughly one twentieth.
A million tokens sounds like a lot, but it's quickly put in perspective: a token is roughly a syllable. So a million output tokens is roughly half a million words — about five novels. MiniMax produces that much text for just over a dollar. With the US flagship, you'd pay over twenty.
And the quality? Honestly: MiniMax isn't playing in the top league. In my tests, Claude is noticeably better on tricky tasks. But not as much better as the price gap would suggest. On the cost-to-benefit ratio, MiniMax absolutely earns its place.
The One Catch — and Where It Still Shines
There is one catch: every now and then — probably when an internal translation goes sideways — MiniMax suddenly drops Chinese characters into the middle of an English text. Charming, but not something I can hand to a client. So everything that goes out professionally runs through Claude. There, quality beats quantity, and the subscription is worth the money.
Where MiniMax stays firmly in use for me anyway: code review. Everything I program first runs through automated tests. Once those are green, an AI model gets to look it over and give me feedback — and that's exactly the job MiniMax M3 does. Here, a slightly clunky phrasing doesn't matter: I know where to look, and I can always ask again.
Put bluntly: I give the instruction, the American does the work, and the Chinese keeps an eye on him. Sounds absurd. Works surprisingly well.
Subscription or Not?
Do I pay for a MiniMax subscription for this? No. I don't come anywhere near the usage at which it pays off — my MiniMax costs average two to three euros a month.
It all runs through OpenRouter, a service that bundles hundreds of AI models behind a single access point. OpenRouter takes about a five percent surcharge for that — which I happily pay. In return, I have one account, one invoice, one access key, instead of a tangle of accounts and subscriptions at every provider separately.
For anyone who needs more, a real subscription can be worth it. Entry costs — like the competition — around $20 a month. The difference: the free allowance is downright generous for the same price compared to Claude or ChatGPT. Credit where it's due. Like Claude's $20 plan, MiniMax bills in rolling windows (one limit every few hours, plus a weekly cap) — you just get considerably further before you hit the wall.
Who Is This For?
It gets interesting mainly for two groups.
First: anyone who burns through tokens — for instance with autonomous AI agents like OpenClaw that chug away in the background for hours. (Anyone running agents like that is probably under-challenged by a beginner's blog anyway — but no matter.)
Second: anyone who keeps slamming into the usage limits of their $20 ChatGPT or Claude plan. Before you reach for the twice-as-expensive US tier, it's worth a quick thought whether a cheap subscription from China serves the same purpose. Also interesting for hobby developers who are sour about GitHub Copilot's new pricing model — but that will probably be an article of its own.
My Verdict
MiniMax belongs in any honest overview of AI models — not as insider-tip hype, but because the models do real work, at a price that makes the US competition look expensive. I use them myself, every day, for code review.
A subscription isn't worth it for me right now: my usage is too low, and for anything important Claude stays the default. But anyone who needs a lot and watches the budget should definitely dare a look east.
Unpaid recommendation, my own opinion — as always here.