MiniMax M3.1 Flash: what it costs and is it free?

MiniMax M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens on the MiniMax API, and cached reads cost $0.015 per million. It isn't free through MiniMax's own API, but MiniMax says the weights arrive on October 23, after which anyone can run it on their own hardware.

Those prices come straight from the launch post. On Monday morning @MiniMax_AI announced the model and, in the same post, cleared up a month-old mystery: "Space Bunny," the anonymous model that had been climbing OpenRouter's usage charts since late September, "was an earlier version of M3.1 Flash."

What MiniMax announced

MiniMax says M3.1 Flash costs a third as much as its previous M3 model, runs three times faster, and scores at the frontier on agent and coding work. Its own write-up puts the cost cut at 70 percent on agentic tasks and typical output speed above 200 tokens per second.

The benchmark numbers in the post:

• DeepSWE (coding): 74%
• Automation-Bench (office "cowork" tasks): 62%
• Agents' Last Exam: 30%
• BabyVision (visual reasoning): 91%

These are MiniMax's own figures, and its blog mostly compares them to unnamed "top models," so I'd give them a few days. Artificial Analysis has already listed the model, though. Its model page scores M3.1 Flash at 42 on its Intelligence Index, fourth of 173 models in its class, and confirms the 1 million token context window and the pricing. It also flags the model as wordy: it used a lot of output tokens in testing. That matters, because output tokens cost four times as much as input.

Where you can use it today

The model is live in the MiniMax API, in MiniMax's M Plan subscription, and in MiniMax Code, its desktop coding agent. MiniMax Code is where it started. On September 27 @MiniMaxAgent said a preview version had debuted there, pitched for work "from quick bug fixes to full features."

Whether it's free depends on where you use it.

Through MiniMax: paid, but cheap

Say a business runs 2,000 customer emails through it for triage, at about 3,000 tokens in and 500 out per email. That comes to roughly $1.50. Both the input and output prices sit below the class medians Artificial Analysis lists.

On OpenRouter: a free tier is listed

OpenRouter lists a free minimax-m3.1-flash:free variant. Free tiers there usually come with rate limits and can vanish without notice, so use it to test the model. I wouldn't build a workflow on it.

Open weights: promised, not shipped

MiniMax says the weights land October 23. The license hasn't been published yet, and Artificial Analysis still lists the model as proprietary. Until the license is out, nobody can say whether you can use it commercially without strings attached.

The Space Bunny reveal

OpenRouter added Space Bunny on September 23 as a stealth model: no maker named, a million token context, free to use. Developers piled in, and community testers noticed its tokenizer behaved like MiniMax's. MiniMax said nothing until today, when its launch post credited the stealth version with more than 200 trillion tokens processed.

A lot of model launches go this way now. A lab ships anonymously, lets developers hammer it for free, and puts its name on it once the usage numbers look good. The catch sits in OpenRouter's terms for stealth models, which say the provider may keep your prompts and completions. If your team pasted customer data into Space Bunny over the past month, that data went to MiniMax, a Chinese lab, before anyone knew who was on the other end.

My read

The price is what I'd pay attention to. Small, fast models are now good enough on real agent work that cost has stopped being the obstacle for most small business automation. At these rates, a model could read every inbound email, sort every support ticket, and draft every quote follow-up for a dollar or two a day.

What's left are data questions: where the model runs, who keeps the logs, and what the license allows. For anything that touches customer data, I'd wait for the open weights after October 23, check the license, and run it on your own hardware or with a US host.

What this means for your business

Model prices keep falling, but a messy process costs the same as it did last year. If your team still retypes emails into a CRM or chases quotes by hand, the model bill is no longer a reason to wait. New Face Design's free process audit maps where your hours go, which tasks a model like this could take over, and where your data shouldn't leave your hands.