Key points
- MiniMax's M3.1 Flash scores 42 on the Artificial Analysis Intelligence Index, up 13 points from M3, and costs under half as much per task.
- Output is billed at four times the input rate, and cached inputs cost a tenth of the standard input rate. Context window is one million tokens.
- Artificial Analysis flags a 93 percent hallucination rate because the model attempts nearly every question rather than declining when unsure.
- Open weights are planned for October 23; the API is already live and a desktop app called MiniMax Code is available.
Why it matters for marketers: Cheaper reasoning models lower the cost of running agent workflows at scale, but a high hallucination rate means content teams still need human review before publishing.
MiniMax has released M3.1 Flash, a reasoning model that reaches 42 on the Artificial Analysis Intelligence Index. That is 13 points above its predecessor M3, and Artificial Analysis puts the cost per task at about $0.22, less than half the earlier model. MiniMax says it also runs three times faster. The model accepts text and images and handles up to one million tokens of context.
The comparison with Chinese rivals is mixed. Deepseek V4.1 Flash 1017 leads overall with 45 points, while M3.1 Flash ties GLM-5.3-Flash and edges past Gemini 3.8 Flash at 41. MiniMax wins on Humanity's Last Exam with 48 percent versus Deepseek's 43 percent, and answers 56 percent of factual questions correctly on AA-Omniscience versus 43 percent for Deepseek. It is also slightly cheaper per task than Deepseek's $0.26, because it uses fewer output tokens.
For marketers, the signal is price pressure. Lower per-task costs make it more practical to run research, drafting and agent workflows in volume. The caveat is reliability: Artificial Analysis reports a 93 percent hallucination rate, since the model tends to answer every question instead of admitting uncertainty. Teams using it for factual content should keep fact-checking and review steps in place.
The API is live now, and MiniMax plans to publish open weights on October 23, following Deepseek, which is already on Hugging Face. Teams that want to self-host or fine-tune should watch that release, and anyone testing the model should benchmark it on their own tasks rather than relying on leaderboard scores alone.
Sources
This brief is our own summary. Read the full reporting at the source.
