China's MiniMax Cuts AI Agent Costs by 70% — But There's a Catch

MiniMax, a Chinese AI company, has released a new model called M3.1 Flash that it claims can run AI agents (software that can perform tasks on its own) 70% cheaper and nearly three times faster than its previous version. The model is available immediately through MiniMax's API and subscription plans, with prices as low as $0.15 per million input tokens. That's a fraction of what top-tier models cost, which could make automation affordable for small businesses and large enterprises alike.

On paper, the model performs well on specialized tests. It scored slightly higher than a competitor from DeepSeek on a business workflow benchmark and tied with Anthropic's Claude Opus 5 on a financial compliance test. Independent evaluator Artificial Analysis found that the average cost to complete a benchmark task dropped from $0.51 to about $0.22. For companies running thousands of automated tasks, those savings add up quickly.

However, there's a significant catch. Independent assessments reveal that M3.1 Flash often invents false answers rather than admitting it doesn't know something. While it answered over half of factual questions correctly, it had a high tendency to make things up. That's a problem because AI agents are increasingly used for tasks like updating accounting software or writing code, where errors can be costly. Faster generation speeds also don't help when agents are waiting for other tools to respond or recovering from mistakes.

The real cost of cheap AI isn't just the price per token—it's the human time needed to check the work. If an agent runs hundreds of operations with hidden errors, the effort to audit and fix them can wipe out any savings. For now, M3.1 Flash might be best for simple, easily checked tasks like drafting code or sorting data. For more sensitive work, businesses will need to weigh the lower upfront cost against the potential for expensive mistakes.