MiniMax's new M3.1 Flash beats DeepSeek and Opus 5 on a key business benchmark at lower cost — but with a high hallucination rate

Chinese artificial intelligence company MiniMax has released M3.1 Flash, a new AI model designed to handle complex enterprise workflows, software development and computer-use tasks at dramatically lower costs than many competing frontier models — including the popular Chinese rival DeepSeek during peak hours.

MiniMax says its new model operates nearly 3X faster than its predecessor while reducing overall usage costs for AI agents by 70%. We’ll dig more into how teams are cutting context and token costs at VentureBeat’s Agentic Infrastructure Stack Summit in San Francisco on October 28.

But the more interesting evidence comes from independent testing: Artificial Analysis has awarded M3.1 Flash a score of 42 on its Intelligence Index, up 13 points from the previous MiniMax M3, with a per-task evaluation cost less than half that of its predecessor.

That puts M3.1 Flash on par with Z.ai's GLM-5.3 Flash, just ahead of Google's Gemini 3.8 Flash, and within striking distance of DeepSeek V4.1 Flash-1017 — another new, inexpensive Chinese AI model — which scores 45.

There is, however, an important limitation: Artificial Analysis also found that while M3.1 Flash answered more than half of its factual knowledge questions correctly, it demonstrated an exceptionally high tendency to generate incorrect answers rather than acknowledge uncertainty.

The results underscore both the growing competitiveness of inexpensive Chinese AI models and the trade-offs enterprises must consider when deploying them for tasks requiring accuracy, reliability and minimal human supervision.

Independent testing puts MiniMax near DeepSeek for less money

According to Artificial Analysis's evaluation, M3.1 Flash achieves an Intelligence Index score of 42 at its maximum reasoning effort setting, compared with 29 for its predecessor, M3.

That represents a substantial improvement in measured capabilities while also taking the average cost of completing an Intelligence Index task from approximately $0.51 down to just $0.22 — a reduction of nearly 57%.

Its closest competitors include GLM-5.3 Flash, which scores 42 at approximately $0.28 per task; DeepSeek V4.1 Flash-1017, which scores 45 at approximately $0.26 per task; and Gemini 3.8 Flash, which scores 41 at approximately $1.24 per task.

While M3.1 Flash remains behind the most capable frontier systems from Anthropic and OpenAI on the overall index, its combination of relatively strong performance and low operating costs makes it an attractive candidate for businesses running AI systems that execute large numbers of tasks.

Artificial Analysis found particularly significant improvements in complex reasoning. M3.1 Flash scored 48% on Humanity's Last Exam, up from M3's 39%, outperforming GLM-5.3 Flash at 40% and DeepSeek V4.1 Flash-1017 at 43%, while matching Gemini 3.8 Flash.

On CritPt, a physics reasoning benchmark, MiniMax improved from just 4% to 22%, ahead of GLM-5.3 Flash's 15% and Gemini 3.8 Flash's 18%, and narrowly behind DeepSeek's 24%.

The new model also used fewer output tokens during the Intelligence Index evaluation — approximately 189 million, according to Artificial Analysis's announcement, compared with roughly 324 million for GLM-5.3 Flash and 334 million for Gemini 3.8 Flash.

That matters because developers pay for the tokens generated by AI models, including the intermediate reasoning required to complete difficult tasks. A model that reaches comparable results while generating fewer tokens can offer substantial savings beyond differences in advertised API rates.

MiniMax targets enterprise automation, coding and computer use

In its official announcement, MiniMax emphasizes the model's ability to execute extended sequences of actions across business software, coding environments and computer interfaces.

The company reports a score of 62.2 on Automation-Bench v1.0.6, which tests AI agents on business workflows involving applications used for sales, marketing, operations, customer support, finance and human resources.

That narrowly exceeds DeepSeek V4.1 Flash-1017's 60.5 and substantially surpasses the reported results of GPT-5.6 Sol and Claude Opus 5 in MiniMax's comparison.

On BankerToolBench, which evaluates multistep financial workflows involving document consistency, compliance and tool use, M3.1 Flash scores 76.2, tying Claude Opus 5 in the company's results.

Its coding results are similarly competitive. MiniMax reports 73.7% on DeepSWE v1.1, compared with 74% for Claude Opus 5, 73% for GPT-5.6 Sol and 77% for DeepSeek V4.1 Flash-1017.

For computer use, the model scores 61.7 on OSWorld 2.0 and 29.6 on Agents' Last Exam, which evaluate an AI system's ability to operate software and complete complex professional tasks.

Although those figures trail the strongest Anthropic and OpenAI models, MiniMax's accompanying cost comparisons suggest the company can deliver much of their practical capability at substantially lower task costs.

The company also reports a score of 82.2 on CyberGym, a cybersecurity benchmark focused on vulnerability discovery and validation.

These are benchmark results rather than documented production deployments, and the different evaluation settings make direct comparisons imperfect. Nevertheless, they suggest that inexpensive models are becoming increasingly viable for demanding enterprise work that previously required much more expensive alternatives.

A 93% hallucination rate raises questions about reliability

One of the most consequential findings from Artificial Analysis concerns factual accuracy.

On its AA-Omniscience benchmark, M3.1 Flash correctly answered 56% of questions, compared with 43% for DeepSeek V4.1 Flash-1017 and 28% for GLM-5.3 Flash.

However, Artificial Analysis found that MiniMax attempted nearly every question, producing a 93% hallucination rate under the benchmark's definition.

Importantly, that does not mean 93% of all its answers were incorrect. The metric measures the proportion of non-correct responses that were outright wrong rather than partial answers or refusals.

In other words, when M3.1 Flash failed to provide a correct answer, it overwhelmingly chose to give an incorrect one rather than acknowledge that it did not know.

That behavior resulted in an overall AA-Omniscience score of just 16, despite the model's relatively strong factual recall.

For enterprises considering AI agents that autonomously update records, analyze financial documents, research customers or make operational decisions, the distinction is significant. High benchmark intelligence alone does not guarantee that a model will reliably identify its own limitations.

Businesses adopting M3.1 Flash for knowledge-intensive workflows may therefore need additional verification, retrieval systems, guardrails or human review to reduce the risk of confidently incorrect outputs.

API pricing starts at $0.15 per million input tokens, with open weights coming October 23

MiniMax is offering M3.1 Flash through its first-party API, M Plan subscription offering and MiniMax Code desktop application.

The model costs $0.15 per million input tokens and $0.60 per million output tokens, with cached input discounted to just $0.015 per million tokens.

At those rates, a workload consuming one million uncached input tokens and producing one million output tokens would cost $0.75 in model usage charges, excluding other infrastructure or application expenses.

This makes M3.1 Flash among the cheapest frontier AI models in currently available around the world.

The model supports text and image inputs, text outputs and a context window of one million tokens, according to Artificial Analysis.

MiniMax also claims that the model can generate more than 200 tokens per second, with median output speeds consistently exceeding 170 tokens per second.

The company attributes the performance improvements partly to increased sparsity, an architectural approach intended to reduce the amount of computation required to generate responses.

The weights remain proprietary at launch, but MiniMax says it plans to release them on October 23. It has not yet provided the licensing terms for that release, making it premature to determine how broadly enterprises will be able to modify or self-host the model.

In another notable disclosure, MiniMax revealed that an earlier version of M3.1 Flash had previously operated anonymously under the name Space Bunny, processing more than 200 trillion tokens before its identity was publicly disclosed. The company has not provided a detailed breakdown of that usage.

MiniMax says a higher-capability Pro model is also planned, although it has not announced a release date.

For enterprise AI teams, M3.1 Flash offers another indication that strong performance on coding and automation tasks is becoming available at increasingly low prices. Its third-party benchmark results reinforce that proposition, while the hallucination findings highlight why pricing and intelligence scores remain only part of the equation when deciding which AI models can be trusted with real business operations.