MiniMax launches M3.1 Flash to lower the cost of AI agents

Key Takeaways

• MiniMax says M3.1 Flash delivers nearly three times faster online inference and 70% lower usage costs than its predecessor
• Company-reported benchmarks point to gains in software engineering, computer use, enterprise automation, and cybersecurity
• Enterprise buyers will still need independent testing, security controls, and production-level cost analysis before deployment

Speed and price are becoming central battlegrounds in enterprise artificial intelligence, particularly for agents that make repeated model calls while completing a task. MiniMax entered that contest on Monday with M3.1 Flash, an updated AI model designed for agentic workflows, coding, computer use, and cybersecurity.

The Chinese AI startup is pitching M3.1 Flash as a substantial step beyond its predecessor. According to Reuters, MiniMax claims the model offers nearly three times faster online inference while reducing usage costs by 70%. Those figures could attract businesses experimenting with agents but struggling to justify the cost of running them at scale.

An ordinary chatbot might answer after one or two model calls. An agent may plan a task, retrieve information, invoke software tools, inspect the result, revise its approach, and repeat the process. Each step consumes tokens and adds latency. Small differences in model pricing and response time can therefore become much larger differences when multiplied across thousands of workflows.

A model that responds quickly enough to support interactive work, while costing less per workflow, can sometimes be more useful than a stronger model that is slower or more expensive to operate.

MiniMax is reporting a peak output rate above 200 tokens per second and a median rate above 170 tokens per second. M3.1 Flash also offers a one-million-token context window, giving developers room to supply large code repositories, extensive documentation, lengthy interaction histories, or several business records within a single session.

The company’s benchmark results are broad. MiniMax Research reports a score of 62.8% on AutomationBench-AA, which evaluates complex, multi-step enterprise-agent tasks. M3.1 Flash also scored 73.7% on DeepSWE v1.1, a software-engineering benchmark spanning code understanding, implementation, debugging, and testing.

For computer-use scenarios, MiniMax reports 61.7 on OSWorld V2. The company also lists a score of 29.6 on Agents’ Last Exam, which is intended to test models on demanding professional work. These results suggest that MiniMax is aiming beyond conversational assistance and toward systems that can operate software, manipulate information, and carry out longer sequences of actions.

Cybersecurity is another part of the pitch. MiniMax reports an 82% score on CyberGym for vulnerability discovery and validation. This figure remains company-reported and needs independent replication. Security benchmarks can be sensitive to prompting methods, tool access, evaluation environments, and whether a model has encountered related material during training.

For enterprise buyers, these claims do not necessitate an immediate model replacement. Benchmark leadership can shift quickly, and aggregate scores rarely capture an organization’s particular data, applications, languages, compliance requirements, or failure costs.

Instead, M3.1 Flash could broaden the list of models worth testing against DeepSeek’s Flash models, OpenAI’s GPT series, and Anthropic’s Claude. The lower-cost positioning may be especially relevant for coding assistants, support automation, internal research agents, and back-office workflows where models repeatedly call tools.

Tool integration will matter just as much as model performance. The Model Context Protocol, or MCP, is increasingly used to connect AI systems with business data and external tools through a common interface. That can simplify development, but it also expands the operational surface that teams need to monitor. Permissions, tool outputs, audit trails, and prompt-injection defenses all affect whether an agent is suitable for production.

Faster inference can improve more than user experience. It may let an agent attempt additional verification steps without making the workflow feel unresponsive. Lower prices can likewise support fallback models, repeated checks, or human review at decision points. Those design choices may be more valuable than simply processing a higher volume of tasks.

Enterprises evaluating M3.1 Flash can use the NIST AI Risk Management Framework to structure testing around reliability, security, transparency, and governance. Practical pilots should measure complete task success, not only token speed. They should also track tool-call failures, unsupported claims, recovery behavior, access-control violations, and total cost per completed workflow.

MiniMax has put forward an ambitious set of performance and cost claims. If real-world testing supports them, M3.1 Flash could intensify pricing pressure across the agent-model market. For buyers, the sensible next step is narrower: test the model on representative workflows, compare full operating costs, and examine how it behaves when tools fail or instructions conflict.