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

MiniMax has put M3.1 Flash on its API, and the pitch is blunt: near-frontier agentic performance without a premium bill. The model is a multimodal mixture-of-experts system with 428B total parameters and about 23B active, plus a 1,000,000-token context window. It takes text, images and video, then returns text.

This is also not a totally new face. MiniMax says Space Bunny was an earlier build of M3.1 Flash, and that stealth version processed more than 200T tokens. What changed between that and the public release is still not explained in a technical report.

The headline number is the price: $0.15 per 1M input tokens and $0.60 per 1M output tokens, with cache reads at $0.015 per 1M and no long-context surcharge. MiniMax is also making the model available through M Plan and the MiniMax Code desktop app, while weights are due on October 23. Self-hosting hardware is still undisclosed.

Performance, at least by MiniMax’s own tables, is strong in the places it wants to win. Automation-Bench v1.0.6 tops out at 62.2, BankerToolBench ties Opus 5 at 76.2, and LOCA-Bench 256K reaches 60.3. But coding is more mixed. Terminal Bench 4.0 comes in at 38.1, while DeepSWE v1.1 lands at 73.7, just ahead of GPT-5.6 Sol and behind DeepSeek-V4.1-Flash-1017.

The model also comes with always-on reasoning, which developers cannot switch off. They get five effort levels — Low, Medium, High, Xhigh and Max — and MiniMax recommends Max by default. That means integrations built around older M3-style <think> tags will need updating, because reasoning now appears in a separate reasoning_content field. In other words, this is not just a cheaper model. It is a different plumbing job.