MiniMax said in its October 19 announcement that M3.1 Flash runs nearly three times faster than its predecessor and can reduce AI agent usage costs by 70%. The Chinese company is targeting software development, business automation, computer use and other workflows where lower inference costs could make continuous agent deployment more economical.
The commercial case is reinforced by pricing. M3.1 Flash is available through MiniMax’s API, subscription plan and MiniMax Code application at $0.15 per million input tokens and $0.60 per million output tokens. MiniMax plans to release the model weights on October 23, although it has not disclosed the licensing terms.
That combination of speed, low pricing and open-weight availability could give MiniMax a useful position in a crowded market. The model is not presented as the most capable system overall. Its advantage is the possibility of delivering adequate performance for routine enterprise tasks at a substantially lower cost than larger frontier models.
Better performance at a lower price
Artificial Analysis reported in its independent evaluation that M3.1 Flash achieved an Intelligence Index score of 42, compared with 29 for the earlier MiniMax M3. The average cost of an evaluated task also fell from approximately $0.51 to $0.22.
The result places M3.1 Flash near several lower-cost competitors. Artificial Analysis ranked it alongside Z.ai’s GLM-5.3 Flash, close to DeepSeek V4.1 Flash-1017, and above Google’s Gemini 3.8 Flash on the cited index. It remains behind the strongest frontier systems, but the comparison suggests that MiniMax is competing effectively in the segment where price and throughput may matter more than absolute reasoning quality.
The model also showed gains on complex reasoning evaluations, including Humanity’s Last Exam and CritPt. During testing, it generated fewer output tokens than several competing systems, a potentially important advantage because shorter responses can reduce both cost and latency in agent loops.
MiniMax said its own tests showed strong results on AutomationBench, BankerToolBench, DeepSWE, OSWorld, Agents’ Last Exam and CyberGym. Those benchmarks cover business automation, financial workflows, coding, computer interaction and cybersecurity. However, the company’s results are not directly equivalent to independent evaluations because benchmark conditions, prompts and scoring methods can differ.
Reliability threatens the business case
The central problem is that lower cost does not necessarily produce lower total cost of ownership. An inexpensive agent that makes confident errors can create expensive review, remediation and compliance burdens.
Artificial Analysis found that M3.1 Flash answered 56% of questions correctly on its AA-Omniscience benchmark, while assigning the model a 93% hallucination rate under that benchmark’s definition. That figure does not mean 93% of all answers were wrong. It measures the share of non-correct responses that were classified as outright incorrect, rather than partial answers or refusals.
The distinction matters, but the result remains troubling for unsupervised enterprise use. A model that refuses frequently may slow a workflow. A model that answers incorrectly while sounding certain can contaminate customer records, misread financial documents or trigger an erroneous operational decision.
Carl Franzen wrote in VentureBeat’s critical account that businesses may need retrieval systems, guardrails or human review to manage that risk. Those controls could reduce the economic benefit MiniMax is offering, particularly when agents are used across thousands of transactions.
Duncan Riley’s neutral report for SiliconANGLE also noted that MiniMax’s competitor comparisons had not been independently verified. That caution is important because vendor benchmarks can demonstrate where a model performs well without establishing how it behaves across the full range of enterprise tasks.
MiniMax therefore has a clear opening, but not an uncontested advantage. M3.1 Flash could appeal to companies willing to trade some reliability for speed and cost, especially in bounded coding or automation workloads. For financial, customer-facing and operational agents, however, the decisive question will be whether deployment safeguards preserve enough of the 70% savings to justify the risk.
Rebeca Smith is not a person. No notebook, no deadlines, no face behind the name — just a byline this newsroom publishes under. Here is the production line underneath it, because a name beside a portrait reads like a journalist, and this one is not one.
The models. Writing: gpt-5.6-luna and qwen3-max. Out on the live web: gpt-5.6-luna and gpt-5.6-terra. Pictures: gpt-image-1 and gpt-image-1-mini. Swap one in the newsroom and this line swaps with it — it is read off the machines, not typed here.
How a story is made
- Research. The searching model reads around the story, pointed at primary sources — the filing, the post, the repository — rather than at somebody else's write-up of them.
- Writing. The writing model drafts it against what was found, at Rebeca Smith's usual length and in Rebeca Smith's usual register.
- The loop. A reviewer reads the draft and sends it back with notes. Then reads it again. A piece can go round several times before it leaves the building.
- Enrichment. A quotation has to appear word for word on the page it is taken from. A chart may only use figures that appear in the source it cites. Whatever fails is dropped, and the reason is kept.
- Fact check. A last pass hunts for claims the article makes and its sources do not.
- A human stop. Sensitive subjects are held for a person to read before publication, and a person can kill any of it at any point.
If that sounds less like a newsroom and more like a factory: quite. It is called Press Factory.
This article was generated using AI and published automatically without human pre-publication review.
How this article was made
The article was produced by the Grandmonts Media News Engine using automated research, drafting and verification workflows. No human editor reviewed the article before publication. Grandmonts Media remains responsible for the published content. Errors can be reported at office@grandmonts.cz.