MiniMax has officially released M3.1 Flash. Before the official announcement, the model ran online under the anonymous codename Space Bunny for more than 200 trillion tokens. Real-world feedback was used to thoroughly strengthen its core capabilities in enterprise workflows, coding, computer control, and cybersecurity offense and defense.
Many models post high benchmark scores but stumble as soon as they enter real business environments. M3.1 Flash took a different path: it was honed on live traffic for long-horizon execution before being put through evaluations.
| Evaluation dimension | Core benchmark | M3.1 Flash | Opus 5 | GPT-5.6 Sol | DeepSeek-V4.1-Flash-1017 |
|---|---|---|---|---|---|
| Enterprise automation | Automation-Bench v1.0.6 | 62.8 | 50.3 | 45.8 | 60.5 |
| Financial tool calling | BankerToolBench | 76.2 | 76.2 | 74.6 | - |
| Long-context control | LOCA-Bench 256K | 60.3 | 53.7 | 60.0 | - |
| Software engineering | DeepSWE v1.1 | 73.7 | 74.0 | 73.0 | 77.0 |
| Desktop cross-application operation | OSWorld 2.0 | 61.7 | 70.6 | 65.7 | - |
| Professional task delivery | Agents' Last Exam | 29.6 | 32.2 | 30.6 | - |
| Cybersecurity offense and defense | CyberGym | 82.2 | - | 83.6 | - |
As the table shows, M3.1 Flash scored 62.8 on Automation-Bench, which tests real-world cross-SaaS business workflows, directly outperforming Opus 5 and GPT-5.6 Sol.
On BankerToolBench, which tests multi-step planning and financial compliance, it also achieved the top score of 76.2.
To support high-frequency calls for long-horizon tasks, MiniMax applied more aggressive sparsification to the model. Online inference is nearly three times faster than the previous generation, everyday output can exceed 200 TPS, median speed remains above 170 TPS, and overall cost has been cut by 70%.
| Billing item | M3.1 Flash price |
|---|---|
| Input | $0.15 / million tokens |
| Output | $0.60 / million tokens |
| Cached input | $0.015 / million tokens |
At this price, completing all 152 complex professional tasks in Agents' Last Exam costs only $111 in total, whereas comparable overseas flagship models often cost thousands of dollars.
The model is now available via API, M Plan, and the MiniMax Code desktop coding tool. Later, MiniMax will open-source the model weights and launch a Pro version aimed at pushing the upper limit of intelligence.