China’s Open-Source Models Challenge Anthropic in AI Coding

Art by Mike Sullivan

Chinese companies are launching open-source AI models built to power coding assistants as cheaper alternatives to those available from OpenAI, Anthropic and Google.

Late last month, Shanghai-based AI startup MiniMax launched M2, a new open-source large language model developed primarily for coding and AI agents. In a post on X, MiniMax noted that the price of M2 is just 8% of that of Anthropic’s Claude Sonnet model. (M2 costs $1.20 per a million output tokens, compared with $15 for Claude Sonnet, according to both companies.)

M2 has already gotten popular. Startups that develop coding agents, such as Cline, Kilo Code and Roo Code, have made M2 available to their users. M2 is currently the fourth most popular model based on token usage on OpenRouter, which offers hundreds of models to developers. M2 trails only xAI, Anthropic and Google on the site.

The low cost isn’t M2’s only selling point. It ranks No. 4 for web development, according to LMArena, which compares AI models on user votes. And it ranks No. 5 in terms of intelligence on another AI model leaderboard, compiled by benchmarking company Artificial Analysis.