MiniMax M2.5: Budget SOTA Model Steals the Show
Chinese startup's new open models slash AI agent costs to $10k/year for continuous operation.
Deep Dive
Chinese AI startup MiniMax released two new open language models, M2.5 and M2.5 Lightning. Built on a Mixture of Experts (MoE) architecture, they achieve near state-of-the-art performance on coding and agentic benchmarks while costing just 1/20th of models like Claude Opus 4.6. The company claims enterprises can run multiple autonomous AI agents continuously for approximately $10,000 per year, using them internally for code generation and operations.
Why It Matters
A 10-20x cost reduction for high-performance AI enables mass deployment of autonomous agents for analytics, content, and workflow automation.