Quick disclosure: Top Tool Stack runs affiliate links on some of the tools we cover. MiniMax is not one of them, there is no program to join and nobody is paying me to like or loathe this thing. Which makes the next part easier to say plainly.
A Shanghai lab most of the room had not thought about until last Tuesday just priced frontier AI like it owes you money.
MiniMax M3 landed on June 1 at thirty cents per million input tokens. And the internet did the thing it does every single week: the collective gasp, the “everything just changed” threads, the pricing-page screenshot passed around like contraband.
I want to give you the calm version, because you are the one who has to explain the bill to someone far less excitable than a quote-tweet.
The one number that is actually real
Strip away the noise and the price is the story. It is a genuine outlier, not a rounding error.
Here is what a million tokens costs you, give or take, right now:
| Model | Input ($/M) | Output ($/M) |
|---|---|---|
| MiniMax M3 (launch promo) | $0.30 | $1.20 |
| MiniMax M3 (regular) | $0.60 | $2.40 |
| Gemini 3.5 Pro | $1.50 | $9.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
| GPT-5.5 | $5.00 | $30.00 |
Read that bottom row again, then the top one. On output tokens, the part that actually piles up when you are running agents in a loop, M3 at promo pricing is roughly one twenty-fifth the cost of GPT-5.5. Even at full freight it is a tenth of Opus.
It also carries a one-million-token context window and takes images and video as input, on a new attention trick MiniMax calls MSA that keeps long-context inference from melting your invoice. If your workload is long documents or browse-and-code agents, that combination at that price is not nothing. For the person paying the bill, this is the most interesting thing to happen to the pricing page all quarter.
So far, so good. Now the part the launch thread skipped.
The asterisks nobody screenshotted
That $0.30 is a “launch promo.” Fifty percent off. The oldest move in the catalogue, the one every gym and broadband company has run since the dawn of the free-trial. Regular pricing is double, at $0.60 and $2.40. Still cheap, still a story, but write the real number on the whiteboard, not the honeymoon one.
Then there is “open-weight,” the phrase doing the heavy lifting in every headline. MiniMax says the weights and the technical report are coming to Hugging Face within roughly ten days of launch. Coming. As in, on launch day the thing everyone was calling open was a closed API with a promise attached. The promise may well be kept (the M2.7 weights are already up there, so the intent looks real), but “open-weight” and “we will open the weights soon, pinky swear” are not the same product, and you should not architect around the second one as if it were the first.
And the benchmarks. M3 posts 59% on SWE-Bench Pro, a hair above GPT-5.5, with a chunky 83.5 on agentic browsing. Lovely numbers. They are also, almost entirely, MiniMax’s own numbers, from MiniMax’s own launch blog. Independent reruns are still thin on the ground, and the early hands-on reports are the usual “mixed in the real world” you get whenever marketing-grade scores meet an actual codebase. Mark these to vendor until someone with no skin in the game reruns them.
One number MiniMax did not put on the hero image: on ARC-AGI-2, the test for novel, you-cannot-memorise-this reasoning, M3 sits under twelve percent. That is in line with the other Chinese frontier models and well behind the US labs. Cheap and fast and long-context, yes. A brain that reasons its way through truly new problems, not so much. Worth knowing before you hand it your hardest thinking.
Who actually gets hurt here
Price wars are great copy and miserable to live through, and the casualties are never the people you would expect.
It is not the Shanghai lab, which has nothing to lose and a reputation to win. It is not OpenAI or Anthropic, who can absorb a quarter of margin compression in their sleep and will simply point at ARC-AGI and their enterprise contracts. It is the thin layer in between. The wrapper startup whose entire pitch was “we give you cheap access to a good model,” watching its only moat get priced to the floor by the model maker itself. The solo founder who built a tidy little margin reselling tokens, now explaining to a board why the spreadsheet changed overnight. The engineer who gets pinged on a Sunday with a link and a one-line instruction to “migrate us to this by Friday, it is a tenth of the price,” as if swapping the brain of a production system were the same as switching electricity providers.
Those are the sacrificial lambs at the altar of AI eFfIcIeNcY gAiNz. The attack here is not on the cheap model, cheap is good, cheap is the whole point of competition. The attack is on the reflex that treats a launch-week promo and a vendor benchmark as a reason to rip out a working stack by Friday. That reflex has a body count, and it is usually the most capable person in the room who gets handed the cleanup.
What I would actually do on Monday
Not panic-migrate. I am a drummer, I know the difference between a loud fill and a song, and most of this week is fill.
If your costs are really killing you, run M3 on your own evals, not MiniMax’s. Your tasks, your data, your definition of “good enough.” A tenth of the price is only a saving if the output does not quietly get worse in ways your users feel before they can name (and they always feel it first).
If your costs are fine, the smartest use of this launch is not a migration at all. It is leverage. A frontier-ish model at a tenth of the price is the best negotiating email you will send your current provider all year. Funny how the rep gets flexible once a real alternative exists.
And if you are still working out which model fits which job in the first place, start with the boring fundamentals before the launch-of-the-week. Our LLM comparison for business and the tools that actually save you ten hours a week will do more for you than any pricing-page screenshot.
The cheapest model in the room is a real event. The weekly insistence that it changes everything is just the noise tax we all pay now. Nail the difference between the two and you are already ahead of most of your market.
So before you forward that pricing page to your engineer with “thoughts?”, one question: are you switching because the model is better for your actual work, or because the number was small enough to feel like a decision? Probably worth answering that before Friday…
Frequently asked questions
What is MiniMax M3?
MiniMax M3 is a large language model from the Shanghai AI lab MiniMax, launched on June 1, 2026. It pairs frontier-level coding performance with a one-million-token context window and native image and video input, on a new sparse-attention architecture (MSA) built to keep long-context inference cheap.
How much does MiniMax M3 cost?
During its launch promotion, M3 is priced at $0.30 per million input tokens and $1.20 per million output tokens on OpenRouter, a fifty percent discount. Regular pricing is $0.60 and $2.40. For comparison, GPT-5.5 runs $5.00 / $30.00 and Claude Opus 4.8 runs $5.00 / $25.00.
Is MiniMax M3 open-weight?
Not on launch day. MiniMax committed to publishing the weights and technical report on Hugging Face within roughly ten days of the June 1 announcement. Until you see the checkpoint listed, treat M3 as an API-only model with open weights promised, not delivered.
Is MiniMax M3 better than GPT-5.5 or Claude Opus 4.8?
On price, comfortably. On vendor-reported coding benchmarks it is competitive with GPT-5.5 and a few points behind Claude Opus 4.8. On abstract-reasoning tests like ARC-AGI-2 it lags the US labs badly. The benchmarks are mostly MiniMax’s own so far, so run your own evaluation before betting a production system on them.
Should I switch my stack to MiniMax M3?
Only after testing it on your own tasks and data, and only once you are reading the regular price rather than the promo. If your current costs are manageable, the better use of M3 is as negotiating leverage with your existing provider.
One short, honest email a week: what’s real, what’s hype, and what’s worth your money. Subscribe and get my full tool-stack guide, free.
Get the free guide