MiniMax M3.1 Flash Review: Agentic Performance, Pricing and Limitations Explained

MiniMax M3.1 Flash delivers fast inference for agentic tasks at half the cost of its predecessor. The model, released October 19, 2026, by Chinese company MiniMax, targets coding, computer use, and cybersecurity workloads with a 1M-token context window and five tunable reasoning levels.

Early tests show competitive results on automation benchmarks alongside a documented high hallucination rate on factual recall. Open weights became available four days after launch under a community license.

MiniMax M3.1 Flash Overview and Key Features

MiniMax M3.1 Flash uses a 428B-parameter mixture-of-experts architecture with roughly 23B active parameters per token. It accepts text, image, and video inputs and generates up to 524K output tokens. Reasoning runs continuously with adjustable effort settings, defaulting to maximum.

The model supports agent loops through sparse attention optimizations that MiniMax claims yield three times faster inference than the June 2026 M3 release. It previously appeared in anonymous testing under the name Space Bunny. Access occurs through the official MiniMax API, M Plan, MiniMax Code, or self-hosted inference via vLLM and SGLang after the October 23, 2026 open-weight release on Hugging Face and ModelScope.

Reported agentic scores include 62.8 on Automation-Bench v1.0.6, 76.2 on BankerToolBench, 73.7 percent on DeepSWE v1.1, 61.7 on OSWorld V2, and 82 percent on CyberGym. The Artificial Analysis Intelligence Index places it at 42–43, up from 29 for the prior version, at an estimated $0.22 per task.

MiniMax M3.1 Flash Pricing, Benchmarks and Comparisons

MiniMax API charges $0.15 per million input tokens, $0.60 per million output tokens, and $0.015 per million cached input tokens with no long-context surcharge. Rates sit roughly half those of M3 and align with DeepSeek V4.1 Flash off-peak pricing; some third-party hosts list marginally lower figures.

Performance comparisons show narrow wins over DeepSeek V4.1 Flash-1017 on Automation-Bench yet trailing results on most coding and terminal benchmarks. The model remains slower and more expensive per task than DeepSeek in several third-party evaluations and falls behind Claude Opus 5 on long-horizon terminal work.

Independent verification remains limited; most published numbers originate from MiniMax. Real-world reliability for unsupervised factual or high-stakes agent tasks requires additional verification layers.

MiniMax M3.1 Flash Limitations and Final Verdict

Artificial Analysis recorded 56 percent accuracy on AA-Omniscience factual recall and a 93 percent hallucination rate when the model produced confident but incorrect answers instead of refusals. Always-on thinking adds latency and cost for simple queries. Self-reported benchmarks dominate available data, and politically sensitive topics lack independent testing.

Early access ended on the free OpenRouter tier October 22, 2026. A higher-capability Pro variant remains expected later. Users running high-volume agent loops gain from the 1M context and low flat rates, yet must implement output checks to offset the documented hallucination tendency.