MiniMax Group Inc released M3.1 Flash on October 19, a 428-billion-parameter mixture-of-experts reasoning model priced at $0.15 per million input tokens and roughly $0.22 per task, undercutting DeepSeek-V4.1-Flash at about $0.26 per task1,2,8. Per-token prices fell by half from MiniMax M3, but The New Stack cautioned that the company's headline claim of 70% agent-cost savings depends on how many tokens an agent spends to finish a job, a figure MiniMax has not disclosed. Artificial Analysis independently measured a 57% reduction in cost per Intelligence Index task, from $0.51 on MiniMax M3 to $0.22 on M3.1 Flash, below MiniMax Group Inc's 70% claim, though the two figures come from different workloads.
The model scores 42 on the Artificial Analysis Intelligence Index, 13 points above MiniMax M3, tying GLM-5.3-Flash and edging past Gemini 3.8 Flash at 41, while trailing DeepSeek-V4.1-Flash at 45.
Agent economics under scrutiny
MiniMax Group Inc said agentic tasks cost 70% less to run on M3.1 Flash than on MiniMax M3, with inference close to three times as fast and output topping 200 tokens per second at everyday peak3. The company attributed the speed gain to more aggressive sparsity at inference time, with the parameter count barely changed from MiniMax M3. Reasoning stays on for every request, with developers choosing one of five effort levels from Low to Max. A study cited by The New Stack found that repeated runs of the same agentic coding task can consume up to 30 times as many tokens, with input tokens accounting for most of the usage, complicating any flat savings claim.
Output costs 60 cents per million tokens, with cached input at 1.5 cents. Jefferies maintained a Buy rating and HK$533 (approximately $68) target price on Minimax-W (00100.HK), saying the model could redefine the capability and cost profile of AI agents6.
Benchmark gains and a hallucination problem
On Humanity's Last Exam, M3.1 Flash hit 48%, beating both DeepSeek-V4.1-Flash and GLM-5.3-Flash. On CritPt, a research-level physics benchmark spanning 71 problems, it scored 22%, ahead of GLM-5.3-Flash at 15% and Gemini 3.8 Flash at 18%, and just behind DeepSeek-V4.1-Flash at 24%. On the AA-Omniscience factual accuracy test, M3.1 Flash answered 56% of questions correctly, well above GLM-5.3-Flash at 28% and DeepSeek-V4.1-Flash at 43%.
MiniMax Group Inc said M3.1 Flash scored 61.7 on OSWorld V2, a computer-operation benchmark, and that no Chinese model has done better. BankerToolBench, a banking workflow test, gave it 76.2, and the company pointed to an 82% score on CyberGym for security work. It scored 73.7% on DeepSWE v1.1, a software-development evaluation5.
The tradeoff is reliability. Artificial Analysis measured a 93% hallucination rate, driven by the model's tendency to answer every test question rather than decline when uncertain.
Availability and what follows
M3.1 Flash accepts text, image, and video input and handles prompts up to one million tokens. The API is live, and the model is also available through the M Plan subscription and MiniMax Code. MiniMax Group Inc said it plans to release M3.1 Flash as open weights on October 23 and that a more capable Pro model is also coming.
MiniMax Group Inc confirmed that Space Bunny, a stealth model that ran on OpenRouter, was an early M3.1 Flash build[1]. Prior Vector Wire coverage noted Space Bunny was OpenRouter's top model for 10 straight days and processed 56.7 trillion tokens on the platform[1].
ANALYSIS The 93% hallucination rate complicates the cost story: agentic workflows that chain dozens of calls amplify each unreliable output, creating a direct tradeoff between per-task price and answer trustworthiness for developers evaluating M3.1 Flash in production agent pipelines.