MiniMax Launches M3.1 Flash, Cuts AI Agent Costs by 70%

MiniMax has released its M3.1 Flash model, positioning it as a significant advancement in both the capability and cost-efficiency of AI agents for practical tasks. The announcement, made on the company's blog, claims the new model delivers a "generational leap" in agentic performance while dramatically reducing the expense of running such systems.

According to the official release, the model is an evolution of a system previously known anonymously as "Space Bunny," which processed over 200 trillion tokens. MiniMax states that M3.1 Flash makes online inference nearly three times faster and reduces overall usage costs for agentic tasks by 70% compared to its predecessor.

What are the key capabilities of MiniMax M3.1 Flash?

The model is designed for complex, multi-step agentic workflows. MiniMax reports that M3.1 Flash ranks among top performers on AutomationBench-AA, a benchmark for completing real enterprise workflows across software-as-a-service (SaaS) systems in domains like sales, marketing, and finance. It scored 76.2 on BankerToolBench, which tests complex planning, document consistency, and chained tool calls.

For software development, it achieves 73.7% on DeepSWE v1.1, a benchmark for complex software engineering tasks that require independent feature development, bug fixing, and testing. In computer use—the ability to operate software and complete cross-application tasks—it scores 61.7 on OSWorld V2 and 29.6 on the Agents’ Last Exam (ALE). The company also claims leading cybersecurity capabilities, with an 82% score on the CyberGym benchmark for vulnerability discovery and validation.

How fast and affordable is the new model?

Speed and cost are central to the M3.1 Flash proposition. MiniMax states that through "more aggressive sparsity" optimizations, the model can reach output speeds of over 200 tokens per second (TPS) in everyday use, with median speeds consistently above 170 TPS.

The pricing model is straightforward. According to the blog post, it costs $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. MiniMax asserts this flat-rate pricing makes the model's overall cost for agentic tasks 70% lower than the previous version, aiming to set a "highly competitive new standard for performance and cost." ## Where does this fit in the current AI landscape? The release underscores the intense competition in producing capable yet cost-effective models for automation. The focus on benchmarks for real-world computer use (OSWorld, ALE), complex coding (DeepSWE), and enterprise automation (AutomationBench) reflects a market shift beyond simple chat towards AI that can execute tasks within existing digital environments. This aligns with broader industry trends, such as those seen in our coverage of [DeepSeek's agent-focused V4.1-Flash-1017](/news/deepseek-releases-v4-1-flash-1017-boosts-agent-and-vision-benchmarks) and the ongoing debate about [AI agents' impact on software output](/news/study-finds-ai-coding-agents-increase-code-volume-but-not-software-output). MiniMax also indicated that a more capable "Pro" version, focused on pushing intelligence boundaries, will follow later, suggesting a tiered strategy similar to other providers. ## How can creators and developers explore agentic AI in Grafy? While Grafy does not directly host the MiniMax M3.1 Flash model, its multimodal workspace is built for experimenting with and comparing the capabilities of different AI systems. The most direct way to explore agentic task workflows is through [Grafy Chat](/chat). Users can access a wide range of cloud models side-by-side, including many via OpenRouter and other providers, to test coding, planning, and analysis tasks. For a structured approach to complex, multi-step AI processes, [Graph](/graph) allows you to build and visualize branching sequences of prompts and outputs, effectively prototyping agentic reasoning. For those specifically interested in the coding benchmarks cited by MiniMax, [Grafy Code](/code) provides an AI-powered studio to test software development tasks. Researchers can use [Grafy Reader](/reader) to upload and interrogate the technical paper or benchmark reports related to models like M3.1 Flash. Furthermore, [Operator](/operator), Grafy's browsing agent, allows for hands-on testing of AI's ability to use tools and execute tasks on the web. ## What are the immediate implications for users? The primary implication is potential cost savings for teams running automated AI agents at scale for customer service, operations, coding, or security analysis. The benchmark results suggest the model could be a practical option for prototyping and deploying certain automated workflows where cost has been a barrier. However, as with any vendor claim, independent verification of the benchmark results and real-world performance is essential. The model is available now via the MiniMax API, their "M Plan," and their "MiniMax Code" desktop application, with model weights promised for a later release. ## Frequently asked questions ### What is MiniMax M3.1 Flash best used for? According to MiniMax, it excels at complex, multi-step agentic tasks like automating SaaS workflows, software development (coding), operating computer applications, and cybersecurity analysis such as vulnerability discovery and patching. It's designed for execution, not just conversation. ### How much does it cost to use? Pricing is $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens. MiniMax claims this represents a 70% overall cost reduction for agentic tasks compared to its predecessor model.

Is the M3.1 Flash model open source?

No, it is currently a proprietary model available via API and MiniMax's own applications. The company states that the model weights "will be released later," but no specific timeline or license details were provided in the announcement.

---

Source: Hacker News — front page — MiniMax M3.1 Flash