Chinese AI startup MiniMax has unveiled M2.5, a frontier model that achieves 80.2% on SWE-bench Verified while costing just $1 per hour at 100 tokens per second. The model excels in coding, agentic tool use, search, and office productivity tasks, with particularly strong multilingual programming capabilities across 10+ languages. M2.5 completes complex coding tasks 37% faster than its predecessor while using 20% fewer search rounds.

Trained with reinforcement learning across hundreds of thousands of real-world environments, M2.5 demonstrates architectural thinking by planning projects before coding. The model shows SOTA performance in multilingual coding tasks and generalizes well across different coding agent harnesses, outperforming Claude Opus 4.6 on Droid and OpenCode benchmarks. MiniMax collaborated with senior professionals in finance, law, and social sciences to train M2.5 for office work scenarios, achieving a 59% win rate against mainstream models in their internal Cowork Agent evaluation.

MiniMax offers two versions: M2.5-Lightning at 100 tokens per second for $0.30 per million input tokens, and M2.5 at 50 tokens per second for half that price. Both support caching and aim to deliver intelligence "too cheap to meter." The company positions M2.5 as the first frontier model where users don't need to worry about cost, potentially enabling innovative new agentic applications across coding, search, and office productivity.