MiniMax M3.1 Flash: China's New AI Price War Challenger Undercuts US Frontier Models by 10x

MiniMax launched its M3.1 Flash model on October 20, 2026, pushing China's AI price war to a new low with a low-cost agentic model that significantly undercuts US frontier models by approximately tenfold.

Performance and Cost Efficiency

The MiniMax M3.1 Flash model achieved a score of 42 points on the Artificial Analysis Intelligence Index, marking a 13-point improvement over its previous iteration, the M3. A key differentiator for M3.1 Flash is its weighted cost per Intelligence Index task, which stands at approximately $0.22. This pricing strategy positions the model at roughly one-tenth the cost of many frontier US models, signaling a notable shift in the competitive landscape for AI services.

In terms of raw pricing, MiniMax M3.1 Flash is set at $0.15 per million input tokens and $0.60 per million output tokens. Additionally, cached input is priced at $0.015. These figures underscore MiniMax's aggressive approach to making advanced AI capabilities more accessible.

Speed and Multimodal Capabilities

The M3.1 Flash model demonstrates substantial improvements in inference speed, running nearly three times faster than the M3. It achieves output speeds exceeding 200 tokens per second, which is critical for applications requiring rapid response times. The model supports a 1M-token context window and is designed to handle both text and image inputs, making it a multimodal solution for various tasks.

Agentic Benchmarks and Hallucination Rates

MiniMax M3.1 Flash has been evaluated across several agentic benchmarks, showcasing its capabilities in complex problem-solving environments:

• DeepSWE v1.1: 73.7%
• BankerToolBench: 76.2%
• OSWorld V2: 61.7%
• Agents' Last Exam: 29.6%
• CyberGym: 82%

Despite its strong performance in these areas, the model exhibits a notable weakness: a 93% hallucination rate on AA-Omniscience. This high rate is attributed to the model's tendency to provide answers even when uncertain, rather than declining to respond. This characteristic may require developers to implement additional safeguards or prompt engineering strategies to mitigate potential inaccuracies.

Availability and Future Plans

The API for MiniMax M3.1 Flash is currently live. Furthermore, MiniMax has announced plans to release open weights for the M3.1 Flash model on October 23, 2026. This move could further democratize access to advanced AI capabilities and foster broader development within the AI community.

Implications for the AI Market

The introduction of MiniMax M3.1 Flash, with its competitive pricing and improved performance, intensifies the global AI price war. By offering a powerful yet affordable agentic model, MiniMax is directly challenging the market dominance of more expensive frontier AI models. This development could lead to increased pressure on other providers to adjust their pricing models or enhance their offerings to remain competitive. The model's multimodal capabilities and agentic performance benchmarks suggest its potential utility across various applications, from coding assistance to complex decision-making tasks.

Conclusion

MiniMax M3.1 Flash represents a significant entry into the AI landscape, particularly for its aggressive pricing strategy and enhanced performance. While its high hallucination rate on certain benchmarks presents a challenge, its overall capabilities and cost-effectiveness position it as a strong contender in the market. The upcoming release of open weights on October 23, 2026, will be a key event to watch, potentially accelerating its adoption and impact on the broader AI ecosystem.

Sources

• models.dev/models/minimax/MiniMax-M3.1-Flash-Preview.toml at dev · anomalyco/models.dev · GitHub
• minimax-m2 · GitHub Topics · GitHub
• provider: MiniMax M3.1 Flash officially supports both 512K and 1M windows, but magpie exposes neither the option list nor a way to select 1M · Issue #2214 · yetone/magpie · GitHub
• nvidia-nim-minimax · GitHub Topics · GitHub