MiniMax launches cheaper AI model for security tasks

Chinese artificial intelligence developer MiniMax Group Inc. has released M3.1 Flash, an updated large language model designed to lower the expense of running autonomous AI agents. According to SiliconANGLE, the release aims to address the high costs associated with tasks that require numerous model calls to complete a single job.

The new model offers inference speeds nearly three times faster than the previous version, with output reaching 200 tokens per second during peak usage. SiliconANGLE reports that MiniMax claims agentic tasks now cost 70% less to execute. The pricing is set at 15 cents per million input tokens, matching the off-peak rate of competitor DeepSeek’s V4.1-Flash model.

What we know

M3.1 Flash contains 428 billion parameters, a size similar to the original M3 model launched in June. SiliconANGLE notes that the speed improvements come from more aggressive sparsity during inference. The model supports text, image, and video inputs with context windows up to one million tokens. Reasoning capabilities remain active for every request, with developers selecting from five effort levels ranging from Low to Max.

In internal testing, the model achieved a score of 61.7 on OSWorld V2, which measures computer operation skills. SiliconANGLE states that MiniMax claims no Chinese model has performed better on this benchmark. On BankerToolBench, a test for banking workflows, M3.1 Flash scored 76.2.

For security applications, the model scored 82% on CyberGym, a benchmark created by University of California, Berkeley researchers to test if AI agents can replicate real software vulnerabilities. SiliconANGLE reports that MiniMax lists code auditing, penetration testing, and malware analysis as suitable tasks for the model. On the DeepSWE v1.1 software engineering test, it managed 73.7% of tasks, compared to 77% for DeepSeek’s model. Neither result has been independently verified.

The launch also identifies the creator of Space Bunny, an anonymous model that appeared on OpenRouter in late September. SiliconANGLE notes that MiniMax’s shares fell after DeepSeek launched its competing model in September, following MiniMax’s initial public offering in January.

Defender's checklist

• Verify that any AI tools used for code auditing are operating within approved security boundaries.
• Monitor for new benchmarks like CyberGym that measure AI capabilities in vulnerability reproduction.
• Check vendor pricing models to understand cost implications of using AI agents for security tasks.
• Ensure that reasoning levels in AI models are set appropriately for the sensitivity of the data.

The background in plain language

AI agents are software systems that can perform complex tasks by making multiple calls to large language models. Each call incurs a cost, so reducing the number of calls or speeding up processing lowers expenses. Benchmarks like CyberGym help organisations assess whether these models can safely handle security tasks such as identifying vulnerabilities. Teams usually evaluate these tools for accuracy and cost before integrating them into workflows.

Background: AI agents

AI agents are the software entities that perform specific tasks using instructions from large language models. Agentic AI is the architectural approach that grants these agents autonomy to plan, reason and act. You use agents as tools within agentic AI systems to achieve complex, multi-step outcomes without manual oversight at every stage.

Read the full guide: AI Agents vs Agentic AI: The Distinction That Matters

What to do and how to stay safe: MiniMax

• Review current AI agent usage to identify opportunities for cost reduction using more efficient models.
• Evaluate new AI security tools against established benchmarks to verify their claimed capabilities.
• Monitor vendor announcements for updates on model performance and pricing changes.
• Once the vendor provides an update, test the model in a controlled environment before production use.

Step-by-step guide: AI Governance for Small Teams: Practical Controls and Cost

General security guidance from the TechWire Beat newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

What is the CyberGym benchmark?

It is a test created by University of California, Berkeley researchers to see if AI agents can reproduce real software vulnerabilities.

How much does M3.1 Flash cost?

MiniMax charges 15 cents per million input tokens for the M3.1 Flash model.

Is M3.1 Flash faster than the previous model?

Yes, MiniMax states that inference is close to three times as fast as the previous M3 model.