The biggest gains are in complex reasoning. The model hits 48 percent on "Humanity's Last Exam", beating both Deepseek (43 percent) and GLM, and scores 22 percent on CritPt, a benchmark of 71 research-level physics problems, ahead of GLM (15 percent) and Gemini (18 percent) but just behind Deepseek (24 percent). On the AA-Omniscience factual accuracy test it answers 56 percent of questions correctly, well above GLM's 28 percent and Deepseek's 43 percent.

Cost per task comes down to about $0.22 because the model emits far fewer output tokens than its rivals. MiniMax says it runs three times faster than M3. API pricing is $0.15 per million input tokens and $0.60 per million output tokens, with cached inputs at $0.015. It handles up to one million tokens of context and accepts text and images.

There is one clear weak spot: Artificial Analysis measured a 93 percent hallucination rate, because the model attempts every test question rather than declining when uncertain. The API is live at platform.minimax.io, MiniMax also offers a desktop app called MiniMax Code, and open weights are planned for October 23. Deepseek's rival model is already on Hugging Face.