MiniMax M3.1 Flash Slashes AI Costs by 70% but Independent Tests Reveal Accuracy Trade-Offs

MiniMax M3.1 Flash Targets Enterprise Automation with Dramatic Cost Reduction

Chinese AI startup MiniMax has released M3.1 Flash, a new large language model designed to handle complex enterprise workflows, software development, and computer-use tasks at significantly reduced costs1

1

. The company claims the AI model operates nearly 3X faster than its predecessor while cutting AI agent costs by 70%, with inference speeds exceeding 200 tokens per second during peak usage2

2

. This cost reduction matters because AI agents typically require dozens of model calls to complete single tasks, making operational expenses a critical concern for businesses deploying automated systems.

M3.1 Flash features 428 billion total parameters with just 23 billion active ones, priced at $0.15 per million input tokens and $0.60 per million output tokens3

3

. The model accepts text, image, and video input with a 1 million token context window and offers five reasoning effort levels from Low to Max2

2

. MiniMax confirmed the model was previously tested anonymously as "Space Bunny" on OpenRouter's routing service, processing over 200 trillion tokens before its official release2

2

.

Independent Benchmarks Show Competitive Performance Against DeepSeek

According to Artificial Analysis's evaluation, M3.1 Flash achieves an Intelligence Index score of 42 at approximately $0.22 per task, representing a 57% cost reduction from its predecessor M3, which scored 29 at $0.51 per task1

1

. This positions the AI model competitively against DeepSeek V4.1 Flash-1017, which scores 45 at $0.26 per task, and matches GLM-5.3 Flash's score of 42 at $0.28 per task1

1

.

On specific benchmarks, M3.1 Flash demonstrated notable improvements. The model scored 48% on Humanity's Last Exam, surpassing GLM-5.3 Flash at 40% and DeepSeek V4.1 Flash-1017 at 43%1

1

. On CritPt, a physics reasoning benchmark, MiniMax improved from 4% to 22%, ahead of GLM-5.3 Flash's 15% and narrowly behind DeepSeek's 24%1

1

. The model also used approximately 189 million output tokens during evaluation, compared to 324 million for GLM-5.3 Flash and 334 million for Gemini 3.8 Flash, translating to substantial savings since developers pay for generated tokens1

1

.

Strong Performance on Enterprise Workflows and Coding Tasks

MiniMax reports M3.1 Flash scored 62.2 on Automation-Bench v1.0.6, which tests AI models on business workflows involving sales, marketing, operations, customer support, finance, and human resources applications1

1

. This narrowly exceeds DeepSeek V4.1 Flash-1017's 60.5 and substantially surpasses reported results from GPT-5.6 Sol and Claude Opus 51

1

. On BankerToolBench, evaluating multistep financial workflows involving document consistency and compliance, M3.1 Flash scored 76.2, matching Claude Opus 51

1

2

2

.

For software development, MiniMax reports 73.7% on DeepSWE v1.1, compared to 74% for Claude Opus 5, 73% for GPT-5.6 Sol, and 77% for DeepSeek V4.1 Flash-10171

1

2

2

. The model achieved 61.7 on OSWorld V2, testing computer operation capabilities, which MiniMax claims represents the highest score among Chinese AI models2

2

. For cybersecurity applications, the company reported an 82% score on CyberGym, suggesting capabilities for code auditing, penetration testing, and malware analysis2

2

.

Critical Hallucination Rate Raises Accuracy Concerns

Despite competitive performance metrics, Artificial Analysis discovered a significant limitation: while M3.1 Flash answered more than half of factual knowledge questions correctly, it demonstrated an exceptionally high tendency to generate incorrect answers rather than acknowledge uncertainty1

1

. This high hallucination rate represents a critical trade-off enterprises must consider when deploying the model for tasks requiring accuracy, reliability, and minimal human supervision1

1

. The tendency to fabricate plausible-sounding but incorrect responses could prove problematic for business applications where factual precision matters, such as financial analysis, legal document review, or medical information processing.

Advanced Sparse Attention Framework Enables Efficiency Gains

M3.1 Flash achieves its performance improvements through MiniMax's proprietary MiniMax Sparse Attention (MSA) framework, which addresses the KV cache bottleneck that typically slows prefill and decode operations3

3

. Rather than evaluating attention token-by-token across the full 1 million token context window, the system compresses KV cache into blocks and calculates low-cost "block scores" to determine relevance, pulling only the most critical 6-7% of blocks into full compute3

3

. This reduces per-token compute requirements to 1/20th of standard attention costs3

3

.

MiniMax developed custom GPU kernels that optimize how Query vectors fetch distributed KV cache data blocks, achieving a 9X speedup during prefill phase and 15X speedup during decode phase at long contexts3

3

. The architecture implements speculative decoding, where a faster draft pathway predicts subsequent tokens while a separate verification pathway corrects prediction errors in parallel3

3

. Paged attention maps memory into dynamic page blocks using strided memory views, eliminating memory fragmentation entirely3

3

.

Market Implications and Future Trajectory

M3.1 Flash is available through the MiniMax API, M Plan subscription, and MiniMax Code desktop app, with cached input priced at 1.5 cents per million tokens2

2

. MiniMax plans to release the model weights later, allowing broader developer access beyond current services2

2

4

4

. The company also announced a more capable Pro model is coming2

2

.

The launch intensifies competition among Chinese AI firms building faster, cheaper alternatives to Western frontier models. However, questions remain about sustainability, particularly given that both MiniMax and DeepSeek operate with compute capacity below 100,000 H100-equivalent GPUs3

3

. With Anthropic and OpenAI clamping down on distillation attempts, uncertainty exists about whether Chinese AI labs can maintain their current trajectory with relatively limited compute resources3

3

. MiniMax, which raised $619 million in its January IPO and trades in Hong Kong, saw shares fall after DeepSeek's original V4.1-Flash launched in September2

2

. Organizations evaluating M3.1 Flash should monitor real-world performance data and conduct thorough accuracy testing before deploying the model for mission-critical automation tasks.