How cheap can AI output get? China: Yes!
MiniMax has released M3.1 Flash, a new reasoning model that scores 42 points on the Artificial Analysis Intelligence Index. That's 13 points higher than its predecessor M3, at less than half the cost per task. The model ties with GLM-5.3-Flash, edges past Gemini 3.8 Flash (41), and trails Deepseek V4.1 Flash 1017 (45).
Complex reasoning saw the biggest improvement. On "Humanity's Last Exam," M3.1 Flash hits 48 percent, beating both Deepseek and GLM. It 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 (15 percent) and Gemini 3.8 Flash (18 percent), and just behind Deepseek (24 percent). For factual accuracy on the AA-Omniscience test, M3.1 Flash answers 56 percent of questions correctly, well above GLM-5.3-Flash (28 percent) and Deepseek (43 percent).
The model also uses far fewer output tokens than its rivals, bringing 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, since the model tries to answer every test question instead of declining when it's uncertain, according to Artificial Analysis.
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, with cached inputs at $0.015. MiniMax plans to release the model as open weights on October 23. The API is live at platform.minimax.io, and 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 and 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, since 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, and MiniMax will follow on October 23.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.