On October 19, 2026, MiniMax officially launched the lightweight large model M3.1Flash, claiming dual breakthroughs in cost and performance: its AA comprehensive intelligence index reached 42 points, up 13 points from the M3 released in June; the average cost per single task is only $0.22, less than half that of the M3 and lower than DeepSeek V4.1 Flash-1017’s $0.26; API pricing is $0.15 per million input tokens and $0.60 per million output tokens, roughly half that of the previous generation. Within 72 hours of launch, it processed over 100 trillion tokens, and the OpenRouter platform recorded 13.7 trillion free calls in seven days, ranking third.
However, the high scores and low prices quickly came under scrutiny. AI review blogger Binkg selected a production management question from AA-Omniscience for a rewriting test: “long-term planning” was changed to “mid-term,” and “number/location/scale of new factories” was replaced with “manpower/equipment/capacity required to complete the master production schedule.” The correct answer should have shifted from RRP to RCCP. Both DeepSeek V4.1 Flash-1017 and Gemini 4 Argon responded to these changes, while M3.1Flash still output RRP. The author replicated this experiment, and the results were consistent; however, after translating the same rewritten question into Chinese and starting a new conversation, M3.1Flash answered correctly again.
Deeper concerns arise from the private leaderboard “Zhizhi,” led by the cat boss: M3.1Flash scores near-perfect marks on retained questions but nearly zero on new ones; moreover, it solves difficult problems with only about 10,000 tokens, whereas the average single question consumes over 40,000 tokens—a consumption pattern that clearly contradicts its reasoning logic. The AA-Omniscience knowledge test further reveals a contradiction: while its accuracy stands at 56.2% (a 39.5-percentage-point jump from M3), its hallucination rate is as high as 92.8%, meaning roughly 41 out of every 100 questions are answered incorrectly—far exceeding M3’s 18.4% and Gemini 3.8 Flash’s 55.2%.
Notably, the anonymous early version of M3.1Flash, “Space Bunny Alpha,” was launched on OpenRouter in September 2026, ranking first in daily API calls for ten consecutive days and processing over 200 trillion tokens in total. On October 23, MiniMax released all model weights as scheduled, allowing researchers to re-run offline evaluations; however, it did not simultaneously make the training data or request logs publicly available, so whether it is “cheating” remains unconfirmed—this could stem from overfitting to the benchmark questions, or from conceptual confusion or a special response mechanism on the interface side. Phenomena such as discrepancies between Chinese and English performance, hallucinations coexisting with high scores, urgently require further independent verification.