MiniMax M3.1 Flash is a cheap, fast AI model that costs $0.15 per million input tokens and $0.60 per million output tokens. It launched on MiniMax’s API on October 19, 2026, and it was the mystery “Space Bunny” model people had been testing.
The price is the headline, but the catch is that most of the praise comes from MiniMax itself. Here’s what’s solid and what still needs proof.
Start with the quick facts.
- Released: October 19, 2026 on the MiniMax API, M Plan, and MiniMax Code
- Price: $0.15 input, $0.60 output, $0.015 for cached input (per million tokens)
- Context window of 1 million tokens, with text, image, and video input
- Thinking: always on, and it can’t be turned off
- Open weights were reported as due October 23, but that isn’t confirmed
What MiniMax M3.1 Flash is and what it costs
MiniMax M3.1 Flash is the final version of the anonymous “Space Bunny” model. MiniMax says that early build handled over 200 trillion tokens before the reveal, per MarkTechPost.
The older MiniMax M3 charges $0.30 input and $1.20 output up to 512K tokens, and the price doubles above that. M3.1 Flash is half that rate with no long-context surcharge.
It reads text, images, and video, but it only writes text. Here’s how the two models compare on price.
| Model | Input (per 1M) | Output (per 1M) |
|---|---|---|
| MiniMax M3.1 Flash | $0.15 | $0.60 |
| MiniMax M3 (up to 512K) | $0.30 | $1.20 |
If you already pay for a bigger model, our look at Mercury 3 pricing shows how fast the low-cost end of the market is moving.
Can MiniMax M3.1 Flash really beat the big models?
On MiniMax’s own charts, yes, for agent-style work. It reports 62.2 on Automation-Bench, ahead of Opus 5 at 50.3 and GPT-5.6 Sol at 45.8. On DeepSWE it scores 73.7.
It’s weaker on some coding tests. MarkTechPost notes Opus 5 leads on Terminal Bench 4.0 (51.8 vs 38.1), and the vendor numbers have no independent replication yet.
Outside testers paint a more mixed picture. Per a Kie.ai deep dive, Artificial Analysis gives it an Intelligence Index of 42 to 43, versus 29 for MiniMax M3. That’s a big jump, but it sits below DeepSeek V4.1 Flash 1017, which scores 45.
Cost per task is where it gets interesting. The same Kie.ai write-up puts M3.1 Flash at about $0.22 per index task, against roughly $0.26 for the DeepSeek model. So it’s cheaper to run, but only by a small margin once reasoning tokens are counted.
In plain terms, the per-token price looks like a huge discount, while the per-task savings are more modest. That gap is why you should test on your own workload.
The catches you should know before switching
Cheap tokens don’t always mean a cheap bill. A few quirks can eat into the savings.
- Thinking is always on, so even simple calls carry extra tokens and delay.
- Artificial Analysis flags a high tendency to give wrong answers instead of saying “I don’t know.”
- Speed claims disagree: MiniMax says 200+ tokens per second, while Artificial Analysis measured about 105.
Turning reasoning off returns an error, and the default effort level is the highest one. Lower the effort setting if you want faster, shorter replies.
The wrong-answer tendency matters most for support bots or research tools. Always keep a human check on anything factual.
How to try it without risking your budget
You can reach the model through MiniMax’s API, which supports OpenAI-style and Anthropic-style calls. There’s also a rate-limited free endpoint on OpenRouter, according to MarkTechPost.
Run your own ten real prompts before moving anything important. Compare cost per finished task, not cost per token, since always-on reasoning changes the math.
Want to see how other new models stack up on price? Our Kimi K3.1 Pro price and benchmarks piece is a good companion.
Frequently Asked Questions
Is MiniMax M3.1 Flash open source?
Not confirmed. Weights were reported as due October 23, but the license hasn’t been stated, so check MiniMax’s own pages before you build on it.
How much does MiniMax M3.1 Flash cost?
List price is $0.15 per million input tokens and $0.60 per million output tokens. Cached input drops to $0.015 per million.
What was Space Bunny?
It was an anonymous early build of this model that MiniMax tested in public before the launch. The final version may behave differently.
Can I turn off thinking mode?
No. Reasoning is always on, and requests that try to disable it return an error. You can only lower the effort level.
Is MiniMax M3.1 Flash worth testing?
Yes, for agent and automation jobs where cost per task matters, because the price is hard to ignore. I’d hold off on trusting the benchmark wins until outside testers confirm them, and I wouldn’t use it for factual answers without a check. Treat it as a strong, cheap worker that still needs supervision.
Leave a Reply