MiniMax launches M3.1 Flash, says usage costs fall 70%
MiniMax announced the official rollout of MiniMax M3.1 Flash. The company says the model that ran anonymously as "Space Bunny" and processed more than 200 trillion tokens was an early version of it.
MiniMax describes M3.1 Flash as a next-generation Flash model that delivers a generational leap in agentic tasks, coding, Computer Use and cybersecurity. It says online inference is nearly three times faster and usage costs are down 70%. The source does not name the predecessor or the baseline for the speed claim. In the pricing section the 70% figure is narrowed to agentic tasks: for those, overall usage cost is 70% lower than the predecessor's.
All the results below are MiniMax's own.
Complex tasks and automation: the company says M3.1 Flash ranks among the top performers on AutomationBench-AA and can complete enterprise workflows across SaaS systems end to end, in sales, marketing, operations, customer service, finance and HR. It scores 76.2 on BankerToolBench (the unit is not stated), which MiniMax ties to multi-step planning, consistency across documents and files, financial compliance and chained tool calls.
Coding: it reaches 73.7% on DeepSWE v1.1, which MiniMax says places it alongside top models. DeepSWE tests whether a model can understand code, build features, fix issues and validate its work through testing on its own. MiniMax says the model holds up over long sequences of steps at a lower cost.
Computer use: it scores 61.7 points on OSWorld V2 (computer operation and cross-application tasks) and 29.6 points on Agents' Last Exam, or ALE (real professional work and complex tasks). MiniMax says it ranks first among Chinese models on both.
Cybersecurity: it scores 82% on CyberGym, which MiniMax presents as evidence of leading ability in vulnerability discovery and validation. MiniMax's internal security experts report strong performance in code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering.
Speed and efficiency: MiniMax credits "more aggressive sparsity" for the faster inference. In everyday use, output speed can reach 200+ tokens per second, and median (P50) speed consistently reaches 170+ TPS.
Pricing is flat per 1M tokens: $0.15 input, $0.60 output and $0.015 cached input. Across AutomationBench, DeepSWE and ALE, MiniMax says the model performs on par with top models at a very low task cost. No specific rival models or per-task costs are named.
M3.1 Flash is available now through the MiniMax API, M Plan and MiniMax Code (a desktop app). Model weights will be released later. A Pro model, which MiniMax says pushes the boundaries of intelligence, will also arrive later. No dates are given for either.
Key facts
- MiniMax rolled out M3.1 Flash, which it says is the release version of the model that ran anonymously as Space Bunny and processed more than 200 trillion tokens.
- MiniMax claims nearly three times faster online inference, with output speeds that can reach 200+ TPS and a median (P50) consistently at 170+ TPS.
- Self-reported scores: 73.7% on DeepSWE v1.1, 61.7 points on OSWorld V2, 29.6 points on ALE, 76.2 on BankerToolBench and 82% on CyberGym.
- Pricing is flat per 1M tokens: $0.15 input, $0.60 output, $0.015 cached input; MiniMax says agentic-task cost is 70% below its predecessor.
- Available now via the MiniMax API, M Plan and MiniMax Code; weights and a Pro model are promised later, with no dates.
Why it matters
This is a low-priced model aimed squarely at agent workloads. MiniMax pairs benchmark results it describes as on par with top models with token prices of $0.15 input and $0.60 output per 1M. It also claims the model ranks first among Chinese models on OSWorld V2 and ALE. The company frames the release as a new standard for performance against cost in agentic tasks. The fact that it ran in public for a while as the anonymous Space Bunny and handled more than 200 trillion tokens is part of the pitch.
Who it affects
Developers and teams building agents, coding assistants and computer-use tools are the obvious audience, since the announcement centres on enterprise workflows, software engineering, computer operation and security work. Anyone running high-volume agent traffic is the target of the cost and speed claims. Security teams are addressed through the CyberGym score and MiniMax's internal evaluations.
How to use it
M3.1 Flash is available through the MiniMax API (platform.minimax.io), the M Plan subscription and MiniMax Code, which has a desktop download. API pricing is flat per 1M tokens: $0.15 input, $0.60 output and $0.015 cached input. Model weights are promised later, so local or self-hosted use is not yet possible; no license is stated.
How solid is it
This is a first-party launch post, and every number in it is MiniMax's own. No independent third-party verification of the benchmark results is mentioned. The comparisons name no rivals: phrases such as "top models" and "leading models" come with no competitor scores. The predecessor is not named, the baseline for "nearly three times faster" is not named, and the unit of the BankerToolBench score is not stated. The security findings come from MiniMax's internal experts.
Risks and caveats
The 70% cost cut is stated against an unnamed predecessor and, in the pricing section, for agentic tasks only, so it should not be read as a cut in list price. The speed figures mix a peak-type number (200+ TPS "can reach") with a median (170+ TPS). The source gives no parameter count, context window or architecture details beyond more aggressive sparsity. The release dates for the weights and for Pro are only "later". The benchmark positioning, such as "ranks first among Chinese models", is a vendor claim.
“Flash represents our ongoing exploration of the balance between efficiency and capability.”
— MiniMax, M3.1 Flash announcement