MiniMax Unveils AI Model M1, Claims 200x Cheaper Training Than GPT-4

MiniMax, a Shanghai-based AI startup, has introduced its latest large language model, M1. The company says the model was trained at a fraction of the cost of top-tier models like OpenAI’s GPT-4.

According to MiniMax, M1 was trained for just $534,700 in compute costs. If true, that’s nearly 200 times cheaper than the estimated $100 million training cost of GPT-4-o, OpenAI’s flagship model.

This claim has not been independently verified. However, the company shared benchmarks that show M1 performing competitively against models from OpenAI, Anthropic, and DeepSeek in terms of both intelligence and creativity.

MiniMax is best known for earlier work involving AI-generated video games. The launch of M1 on June 16, 2025, marks a major shift in its product direction. It also places the startup in the global spotlight alongside other recent challengers like DeepSeek, which released its R1 model in January, and Butterfly Effect, which launched the Manus “agentic AI” model in March.

Fortune noted that M1’s training cost could be a game-changer. As OpenAI’s cost structure continues to balloon, investors are beginning to ask questions. A report from The Information in October 2024 revealed that OpenAI was projected to lose $14 billion in 2026 and may not break even before 2028.

Tensions between OpenAI and Microsoft, its largest investor and partner, have also been growing. If MiniMax’s claims are accurate, the economic advantage could shift dramatically.

GenAI has been criticized for its poor financial performance, with some calling the business model unsustainable. MiniMax’s low-cost approach could change that narrative, especially if M1 delivers strong results in real-world applications.

As of now, no third-party testing has confirmed M1’s performance. But industry watchers are paying close attention. The potential disruption is significant, especially as Western tech giants continue to invest heavily in increasingly expensive models.

With China making moves in efficient model training, many in the AI world are wondering if the global balance of power in GenAI is beginning to shift.