MiniMax releases M3.1 Flash at $0.15 per million tokens
MiniMax has launched M3.1 Flash, a low-cost agentic and coding model that it says matches leading models on several benchmarks while cutting agent-task costs by 70%, with open weights promised later.
MiniMax announced the general release of M3.1 Flash and confirmed that "Space Bunny," a model that processed more than 200 trillion tokens while running anonymously, was an early version of it.
The Chinese lab is targeting agentic work, coding, computer use and cybersecurity. All of the benchmark results below come from MiniMax itself:
- DeepSWE v1.1: 73.7%
- OSWorld V2: 61.7
- Agents' Last Exam (ALE): 29.6. MiniMax says the model ranks first among Chinese models on both ALE and OSWorld V2.
- BankerToolBench: 76.2
- CyberGym: 82%
MiniMax also says the model is among the top performers on AutomationBench-AA. The post does not include scores for competing models.
The company says it used "more aggressive sparsity" to speed up inference. It reports output speeds above 200 tokens per second and a median (P50) above 170 TPS, which it describes as nearly three times faster than before. MiniMax says the overall cost of agentic tasks is 70% lower than with the previous model.
Pricing is flat at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens.
M3.1 Flash is available now through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. The company says it will release the model weights later but has not given a date. A larger Pro model, which MiniMax says "pushes the boundaries of intelligence," is also due later, with no timeline given. The post does not disclose the model's parameter count, context length or architecture.