China's MiniMax launches cheap AI model to rival Anthropic, OpenAI

China’s AI company MiniMax has launched a new language model, M2.5, aimed at real-world productivity, amid a busy week for China’s AI industry.

The company claims M2.5 performs comparably to models from US firms like Anthropic and OpenAI in tasks such as coding and search, based on its internal benchmarks.

The model, with 230 billion parameters, is designed to be cost-efficient, allowing continuous use at a rate of 100 tokens per second for US$1.

MiniMax highlighted its model’s integration into its AI agent product, MiniMax Agent, which autonomously performs office tasks.

The company reported that 30% of internal tasks across various departments were completed by M2.5.

🔗 Source: South China Morning Post

🧠 Food for thought

Implications, context, and why it matters.

Efficiency comes from the architecture, not the sticker price

M2.5 saves money through its Mixture-of-Experts (MoE) architecture, which activates 10 billion of 230 billion total parameters per task 1.
That setup supports pricing of $0.15 per 1M input tokens for standard M2.5, compared with $5.00 per 1M input tokens for Claude Opus 4.6, based on VentureBeat’s cited pricing table 1.
On SWE-Bench Verified (a test that measures how well models fix real software issues), VentureBeat puts M2.5 at 80.2% and says it comes close to Anthropic’s Claude Opus 4.6 1.
The company calls the model “open” or “open source,” yet the weights (the trained model parameters) and code were not posted and the license terms were still undisclosed at the time of reporting 1.

Lower costs push work beyond chatbots and toward autonomous agents

Cheaper inference makes it easier to run AI “agents” that handle multi-step work on their own, similar to MiniMax Agent for office tasks performed on a user’s behalf 1.
MiniMax says M2.5 generates 80% of its newly committed code 1.
With high-performance AI priced lower, more teams can justify workflows that once cost too much, including large-scale code audits or continuous financial analysis 1.
Smaller groups can also run more trial runs of agentic systems, since development and testing can burn through large token volumes 2.