MiniMax has introduced M3.1 Flash, a new version of its AI model that processes over 200 trillion tokens while operating under an anonymous codename.

The model is designed to improve performance in agentic workflows, software development, computer interaction and cybersecurity. According to internal benchmarks, it achieves speeds up to three times faster than previous versions while reducing operational costs by 70%.

M3.1 Flash scores 76.2 on AutomationBench-AA, placing it among leading models for executing multi-step enterprise tasks across sales, marketing, finance and customer service. It also scores 73.7 on DeepSWE v1.1, a benchmark for complex software engineering tasks requiring independent code understanding and validation. In cybersecurity evaluations, the model achieved 82% accuracy on CyberGym, demonstrating capabilities in vulnerability detection and patching.

The model scored 61.7 on OSWorld V2 and 29.6 on Agents’ Last Exam (ALE), ranking first among Chinese-developed models on both. These results reflect its ability to interact with real-world environments, use external tools and maintain consistency across extended task sequences.

Pricing for M3.1 Flash is structured at $0.15 per million input tokens, $0.60 per million output tokens and $0.015 per million cached input tokens. This represents a 70% reduction in cost compared to earlier iterations, with the company noting that task-level expenses are now significantly lower than competing models.

MiniMax states that M3.1 Flash establishes a new benchmark for balancing performance and affordability in AI systems designed for autonomous workflows. The model is accessible through the MiniMax API, M Plan subscription service and MiniMax Code desktop application. Weights for the model are scheduled for release at a later date.

The company describes Flash as part of an ongoing effort to make advanced AI capabilities more widely available. A more advanced variant, referred to as Pro, is planned for future deployment.