MiniMax’s new M3.1 Flash model combines stronger benchmark results with a lower reported cost per task. Its scores put it near competing models from China and elsewhere, though the comparisons show no single model leading every test.
Higher scores at a lower task cost
M3.1 Flash reaches 42 points on the Artificial Analysis Intelligence Index, 13 points above its predecessor, M3. The article reports that the new model costs less than half as much per task as M3.
On the index, M3.1 Flash ties with GLM-5.3-Flash, scores one point above Gemini 3.8 Flash at 41, and remains below Deepseek V4.1 Flash 1017 at 45. These results place it among closely matched models rather than establishing a lead across the board.
The largest improvement is in complex reasoning. On “Humanity’s Last Exam,” M3.1 Flash scores 48 percent, ahead of both Deepseek and GLM. It also earns 22 percent on CritPt, a test of research-level physics reasoning across 71 complex problems. That is above GLM-5.3-Flash at 15 percent and Gemini 3.8 Flash at 18 percent, while Deepseek scores 24 percent.
Different tests favor different models
Factual accuracy provides another point of distinction. On the AA-Omniscience test, M3.1 Flash answers 56 percent of questions correctly. That is higher than Deepseek’s 43 percent and GLM-5.3-Flash’s 28 percent.
In a direct comparison with Deepseek V4.1 Flash 1017, the results are mixed. Deepseek leads the Intelligence Index and CritPt, while MiniMax scores higher on “Humanity’s Last Exam” and AA-Omniscience. The figures suggest that performance depends on which capability a user needs to evaluate.
Cost also favors MiniMax in this comparison. The reported cost per task is about $0.22 for M3.1 Flash, compared with $0.26 for Deepseek. Across the benchmark, MiniMax used about 189 million output tokens, versus 224 million for Deepseek, helping explain the difference.
Efficiency comes with a reliability concern
MiniMax says M3.1 Flash runs three times faster than M3. The model also uses fewer output tokens than its rivals, another factor behind its lower task cost. For people paying by API usage, lower token consumption can matter alongside benchmark scores.
However, Artificial Analysis reports a 93 percent hallucination rate. The stated reason is that the model attempts to answer every test question instead of declining when uncertain. That behavior is a significant consideration for tasks where an unsupported answer could be more costly than no answer.
The benchmark results and the hallucination figure describe different aspects of the model. Strong scores on selected tests do not remove the need to check its answers, particularly when the model is pressed to respond even without confidence.
API access and planned open weights
M3.1 Flash accepts text and images and supports up to one million tokens of context. Its API pricing is $0.15 per million input tokens and $0.60 per million output tokens; cached inputs cost $0.015. The API is live at platform.minimax.io, and MiniMax also offers a desktop app called MiniMax Code.
MiniMax plans to release M3.1 Flash as open weights on October 23. Deepseek’s model is already on Hugging Face, according to the source article. The planned release gives developers another option to assess, while the reported pricing and benchmark differences offer a starting point for comparison.