MiniMax released M3.1 Flash on October 19, and the pricing is the headline: $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million for cached input, flat across a one-million-token context window. That is half of what its predecessor M3 charged under the same 512k-input tier, and MiniMax claims a 70 percent overall cost reduction for agentic workloads. Inference is claimed to run nearly three times faster.
There is a good story behind the release. An early version of the model ran anonymously on OpenRouter under the name Space Bunny and processed more than 200 trillion tokens before MiniMax owned up to it. That means this model has been quietly load-tested by real users at real scale, which is more than most launches can say.
The numbers, both kinds
MiniMax’s own benchmarks are aggressive: 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2, and around 82 percent on CyberGym, with claims of leading placement among Chinese models on agent evaluations. Vendor benchmarks should always be read with one eyebrow raised, and here there is a reason for that.
Artificial Analysis independently scored the model at 42 on its Intelligence Index, up from 29 for M3, at roughly $0.22 per task. That tiers it with Z.ai’s GLM-5.3 Flash, slightly ahead of Gemini 3.8 Flash, and behind DeepSeek V4.1 Flash-1017. VentureBeat’s coverage flags the less flattering independent finding: a 93 percent hallucination rate on the AA-Omniscience benchmark. When the model failed a factual question, it almost always answered confidently and wrongly instead of admitting uncertainty.
The Hacker News thread, which drew over 400 points, was similarly divided. Several commenters who had used Space Bunny in production found it underwhelming next to DeepSeek V4.1 Flash, and called it verbose. Verbose matters more than it sounds: a model that burns extra tokens on reasoning can cost more per task than its sticker price suggests, even at $0.15 input.
The open weights question
The API is live now through MiniMax’s platform, M Plan subscription, and MiniMax Code desktop app. The weights are proprietary at launch, with a promised release on October 23. Crucially, the license terms have not been announced. The predecessor M3 shipped under a custom community license with a $20 million revenue threshold, which is not the same thing as an MIT license, and MiniMax has given no reason to assume the new terms will be looser.
This is the part worth waiting for. DeepSeek’s V4.1 Flash is already out under MIT with weights on Hugging Face, so it wins the self-hosting argument by default until MiniMax publishes its license file. If your team’s plan requires running the model on your own GPUs, calendar the October 23 date and read the license before committing anything.
The Space Bunny play is becoming a pattern
The stealth-launch tactic deserves its own note. Putting a new model on a router under a nonsense codename, letting thousands of developers hit it with real workloads for weeks, then revealing it with 200 trillion tokens of usage stats attached is a smarter evaluation strategy than any press release. It produces distribution-scale evidence about reliability, verbosity, and failure modes before launch day, and it sidesteps the brand halo that makes people grade flagship models generously. Users judged Space Bunny as an anonymous mid-tier model, which is a harder and more honest test. Expect more labs to run this play, and treat it as a buyer’s lesson too: the anonymous model you tried last month may already be the release everyone is discussing today, so it pays to keep notes on the unnamed ones that impressed you.
What this means for buyers
The interesting shift is not this specific model. It is that flagship-adjacent agentic performance now costs $0.75 per million uncached input-plus-output tokens. At that price, the economics of running an agent that reads a hundred pages before answering, or an overnight coding loop that burns through thousands of turns, change from budgeted line item to rounding error. Premium-priced frontier models lose their pricing power on exactly these workloads.
For anyone building high-volume agent pipelines, the practical move is unglamorous: benchmark M3.1 Flash against DeepSeek V4.1 Flash and GLM-5.3 Flash on your own workload. They are all in the same price class now, so the decider is task-level cost and reliability on your data, not anyone’s leaderboard. Keep a hallucination-sensitive workload on a model that knows when to say I do not know, because that is the one axis where the independent testing here was genuinely worrying.
One more practical note: prompt caching reads at $0.015 per million tokens reward agent designs that reuse long system prompts and document contexts, so the cheapest architecture and the fastest architecture increasingly point the same direction. The cheap-model generation keeps arriving faster than anyone can evaluate it. The winning strategy has not changed, though: trust the independent indexes, distrust the launch blog, and let your own traces pick the model.