MiniMax M3.1 Flash Makes Cheap AI Agents More Capable, but Accuracy Still Matters

A cheap AI model can be impressive at finishing a business workflow and still be a poor source of facts. That is the tension in MiniMax’s new M3.1 Flash, released on October 19. The Chinese AI company is pitching it as a faster, lower-cost model for software agents, while independent testing suggests buyers should examine reliability as closely as task scores.

In its launch announcement, MiniMax says M3.1 Flash is designed for multi-step work across business software, coding and computer use. The company claims inference is nearly three times faster and the cost of agentic tasks is 70% lower than with its predecessor. Those are provider claims, not a promise that every customer will see the same improvement.

Artificial Analysis lists M3.1 Flash at 73% on AutomationBench-AA, a test of agents carrying out workflows across software services. Its broader Intelligence Index score is 42. The same evaluator puts the average cost at about $0.22 per task in its testing. Together, those figures make the model a serious option for teams that need to run many routine actions without paying frontier-model prices each time.

But completing a task is not the same as getting every statement right. Artificial Analysis also tracks knowledge and hallucination through AA-Omniscience. Its current index score for M3.1 Flash is 16. That is a distinct measure from AutomationBench-AA and should not be read as a 16% hallucination rate. A useful procurement question is how often the model confidently invents information when it is asked to explain a decision, summarize a document or work with uncertain facts.

This matters more when an agent can take action. A mistaken answer in a chatbot may be caught before it goes anywhere. A mistaken instruction inside an automated finance, customer-support or operations workflow can move through several systems before a person notices it. Business users therefore need tests based on their own documents, permission controls and a human approval step for consequential actions.

MiniMax lists API prices of $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. At those rates, high-volume experimentation becomes easier for startups and enterprises that cannot justify expensive models on every request. The company says M3.1 Flash is available through its API, M Plan and MiniMax Code.

MiniMax also says it will release model weights later, but it has not set a firm date in the launch post. For now, businesses should assess it as an available hosted model, not assume they can already deploy the weights themselves. TechBooky’s guide to open-weight AI explains why that distinction affects cost, control and portability.

The larger significance is competitive. Providers are no longer judged only on whether a model tops a general leaderboard. They must show that it can complete real work quickly, at a sustainable price, without creating an unmanageable verification burden. M3.1 Flash strengthens the case for cheaper AI agents. Its ultimate value will depend on how reliably those agents behave outside a benchmark, where mistakes have customers and costs attached.