Chinese AI firm MiniMax has launched M2.5, a new model designed for real-world productivity. The company claims M2.5 matches top models from US firms Anthropic and OpenAI in coding and search tasks. With 230 billion parameters, M2.5 keeps the same size as previous versions but offers faster and cheaper operation. MiniMax said, "M2.5 is the first frontier model where users do not need to worry about cost, delivering on the promise of intelligence too cheap to meter." The model can run for an hour at 100 tokens per second for just $1. Tokens are small data chunks models process. MiniMax used M2.5 in its AI agent product, MiniMax Agent, acting as an "AI intern" for office work. The new model helped complete 30% of internal tasks autonomously in research, sales, and HR units. On an internal benchmark inspired by OpenAI’s GDPval, M2.5 scored 59%, up from 24.6% for its older M2.1 model. By comparison, the top model, Anthropic’s Claude Opus 4.6, scored 73.5%. MiniMax did not reveal the chips used to train M2.5 but promised to open-source the model’s mathematical parameters. MiniMax’s IPO prospectus notes that over 70% of its revenue comes from overseas markets, led by its AI companion app Talkie and video app Hailuo. The launch follows a busy week of AI model releases in China, including rival Zhipu AI’s GLM-5 with 744 billion parameters. MiniMax remains a key player in China’s fast-growing AI sector.