MiniMax M3.1 Flash scores 42 on the Artificial Analysis Intelligence Index

MiniMax has launched M3.1 Flash, a new reasoning model that costs less than half as much per task as its predecessor, M3.

The model matches GLM-5.3-Flash and edges ahead of Gemini 3.8 Flash, which sits at 41. It trails Deepseek V4.1 Flash 1017, which scored 45. China continues to push the price of AI output down.

Complex reasoning saw the biggest improvement. On the Humanity’s Last Exam benchmark, M3.1 Flash hits 48 percent. It beats both Deepseek and GLM. The model also scores 22 percent on the CritPt benchmark, which tests research-level physics reasoning across 71 complex problems. That puts it ahead of GLM-5.3-Flash at 15 percent and Gemini 3.8 Flash at 18 percent. It sits just behind Deepseek at 24 percent. For factual accuracy on the AA-Omniscience test, M3.1 Flash answers 56 percent of questions correctly. That is well above GLM-5.3-Flash at 28 percent and Deepseek at 43 percent.

The model uses far fewer output tokens than its rivals. This brings the cost per task down to about $0.22. MiniMax says it runs three times faster than M3. One weak spot is its 93 percent hallucination rate. Artificial Analysis notes this occurs because the model tries to answer every test question instead of declining when it is uncertain.

The model handles up to one million tokens of context and accepts both text and images. API pricing sits at $0.15 per million input tokens and $0.60 per million output tokens. Cached inputs cost $0.015. MiniMax plans to release the model as open weights on October 23. The API is live at platform.minimax.io. MiniMax also offers a desktop app called MiniMax Code.

Deepseek and MiniMax split results in China’s latest pricing battle

Compared head-to-head with the recently released Deepseek V4.1 Flash 1017, neither model wins across the board. Deepseek leads on the Intelligence Index with 45 points versus MiniMax’s 42. Deepseek has a slight edge on the CritPt benchmark at 24 percent versus 22 percent.

MiniMax pulls ahead on other tests. It hits 48 percent on Humanity’s Last Exam, while Deepseek manages 43 percent. For factual accuracy on the AA-Omniscience test, MiniMax answers 56 percent of questions correctly versus Deepseek’s 43 percent. MiniMax is also slightly cheaper per task at $0.22 compared to Deepseek’s $0.26. This is because it needs fewer output tokens across the benchmark, about 189 million versus 224 million for Deepseek.

Both models are getting open-weight releases. Deepseek is already on Hugging Face. MiniMax will follow on October 23.