MiniMax M3.1 Flash costs $0.15 per million input tokens and $0.60 per million output tokens, and MiniMax says the model's overall cost on agentic tasks is 70% lower than its predecessor's. MiniMax built the model for agent work such as coding, operating a computer and multi-step business workflows.
Cache hits cost $0.015 per million tokens, one-tenth of the standard input rate, which matters for agents that resend the same long context at every step of a task. MiniMax charges $0.30 per million input tokens and $1.20 per million output for its M3 model on the standard tier with prompts up to 512k tokens, so Flash costs half as much on both.
| Token type | USD per million tokens |
|---|---|
| Input | $0.15 |
| Output | $0.60 |
| Cached input | $0.015 |
The model had a public run under a codename. An early version circulated anonymously as Space Bunny and processed more than 200 trillion tokens, the company says. MiniMax calls online inference nearly three times faster without naming the baseline, and credits more aggressive sparsity for output speeds of 200+ tokens per second in everyday use, with a median above 170.
M3.1 Flash scores 73.7% on DeepSWE v1.1, a test where models must understand code, build features, fix issues and verify their own work. It reaches 61.7 on OSWorld V2 for computer operation and 29.6 on Agents' Last Exam, which covers professional tasks, and 82% on CyberGym for vulnerability discovery.
MiniMax ranks the model first among Chinese models on OSWorld V2 and Agents' Last Exam and places it alongside top models on DeepSWE. Its security claims beyond CyberGym rest on evaluations by its own experts in code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering.
Developers can try it through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Model weights will be released later, and a Pro model aimed at higher intelligence is planned to follow.
