MiniMax Just Launched M3.1 Flash: 428B MoE Agentic Model at $0.15/M Input

MiniMax released M3.1 Flash on its API at $0.15 per million input tokens and $0.60 per million output, with no surcharge across a 1 million-token context. It's a mixture-of-experts model with about 23 billion of its 428 billion parameters active, and it takes text, image and video in. MiniMax also confirmed an earlier stealth model called Space Bunny was a build of this one.

The pitch is near-frontier agent performance for office-style automation at a fraction of the usual price. If that holds up, it pushes costs down again for anyone running agents at volume.

The price is the story we can verify right now. Per token, this is half the standard rate of MiniMax's own M3, and well below rivals listed in the comparison, like Kimi K3 at $3 in and $15 out. MiniMax wants buyers to think in cost per task. On its own Automation-Bench chart, every effort level comes in under five cents per task. That's the right frame for agent workloads, and we're glad more labs are using it.

The benchmarks need a caution label. Every score is MiniMax-reported, with no independent replication yet. The model leads on Automation-Bench and ties Opus 5 on a banking tool benchmark. It's weaker on hard coding: 38.1 on Terminal Bench 4.0 against 51.8 for Opus 5. There's even a small mismatch between MiniMax's chart and table on its best result. Spot checks against another lab's model card lined up, which helps. But harness choices move these numbers a lot.

One design choice deserves attention from developers. Reasoning is always on and can't be disabled, and MiniMax recommends the highest of five effort levels by default. That sets a floor on tokens and latency even for simple calls, and requests that try to turn thinking off will error. Cheap per token isn't always cheap per call.

We're encouraged by the open-weights promise, with weights slated for October 23. A strong, inexpensive model you can eventually run yourself is good for competition and for anyone wary of depending on a single vendor. The open questions are the license, which hasn't been disclosed, and the hardware needed to host something this size. I'm watching both before calling this a self-hosting option.

Chinese labs including MiniMax, DeepSeek, Kimi and GLM have been competing hard on price and open weights. DeepSeek V4.1 Flash, for comparison, posts a higher score on one coding benchmark in MiniMax's own table and ships under an MIT license. MiniMax hasn't published a technical report explaining what changed from M3.