MiniMax-M3 debuts, eclipsing GPT-5.5 and Gemini 3.1 Pro on key benchmark performance for just 5-10% of the cost

Chinese AI startup MiniMax launched its M3 large language model on Sunday, delivering frontier-level coding and agentic performance with a 1-million-token context window and native multimodality at a fraction of competitors' costs. Pricing begins at $20 per month under subscription plans, or $0.3 per 1 million input tokens and $1.20 per 1 million output tokens for the next week—far below the rates of U.S. models like GPT-5.5 or Gemini 3.1 Pro. Even at full price ($0.6/$2.40 per million tokens), M3 remains 8-20% the cost of leading proprietary models. The model disrupts the traditional trade-off between high-performance closed-source AI and cost-effective open alternatives, combining multi-step reasoning, dense coding, and large-scale data handling without sacrificing efficiency. MiniMax plans to release an open-source version with full customization, including 'open weights,' within the next 10 days, though the model is currently accessible via API. At its core, M3 leverages a novel **MiniMax Sparse Attention (MSA)** technique to reduce computational costs by avoiding the quadratic scaling of traditional Transformer networks. Instead of reprocessing entire datasets, MSA pre-filters Key-Value matrices into precise blocks, dynamically aggregating only relevant queries. Internal tests show MSA achieves over 4x faster performance than alternatives by optimizing memory access and hardware utilization. The pricing comparison underscores M3’s competitive edge: while models like GPT-5.5 charge $5 per million input tokens and $30 per million output tokens, M3’s rates remain significantly lower. The company’s approach challenges the dominance of U.S.-based AI giants, offering both enterprise-grade performance and open-source flexibility. MiniMax’s leadership emphasized the model’s dual appeal—enterprise-grade capabilities at a fraction of the cost, alongside planned open-source accessibility. The release marks a shift in AI development paradigms, potentially democratizing advanced model deployment for developers and businesses.