🔍 Read the full analysis: MiniMax M3.1 Flash: The Results From Our Test Against Open Models on ThorstenMeyerAI.com
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
MiniMax released M3.1 Flash on October 19, 2026, pricing it at $0.15 per million input tokens and $0.60 per million output tokens. The source reports a score of 43 on Artificial Analysis’s Intelligence Index and describes a 24-article newsroom test, but the supplied material ends before giving that test’s results. The model’s weights are promised for October 23; its parameter count and license have not been disclosed.
MiniMax says the new model costs half as much per token as M3 and about 70% less per agent task. The task-cost figure is the company’s claim, not an independently verified result in the report. MiniMax also says everyday output speed exceeds 200 tokens per second, with a median above 170, and describes the model’s design only as using “more aggressive sparsity.”
Artificial Analysis measured 120 tokens per second and 1.08 seconds to first token. Its Intelligence Index gave M3.1 Flash a score of 43, behind MiMo-V2.6-Pro at 46, DeepSeek V4.1 Flash 1017 and GLM-5.3 at 45, and Step 5 Preview and Kimi K3 at 44. The index also recorded 180 million tokens used to complete its evaluations, compared with a median of 99 million. The report says this high token use can reduce the practical benefit of a low per-token price.
MiniMax’s own benchmark charts show a mixed comparison with selected competitors. M3.1 Flash scored 62.2 on Automation-Bench, ahead of DeepSeek V4.1 Flash 1017’s 60.5, but DeepSeek scored higher on DeepSWE and Terminal-Bench. MiniMax’s charts do not report every rival’s score on every benchmark, and the company selected the models for comparison. ThorstenMeyerAI.com notes that MiniMax’s security results, including an 82% CyberGym score, come from an internal evaluation and have not been independently reproduced.
Infographic 01 · MiniMax M3.1 Flash
A step behind on the index. Ahead of every open model in our newsroom.
An independent intelligence index puts MiniMax M3.1 Flash a step behind the open-weight leaders. Writing real news from real sources, it beat all of them.
Independent index
Artificial Analysis Intelligence Index, 10 evaluations. Bar length on a 0–50 scale.
Artificial Analysis lists M3.1 Flash as proprietary until its weights are released.
Our newsroom test
Mean score, English articles, 1–5 by GPT-6.1 Sol and Claude Opus 5.5 with the source in context. Bar length on a 0–5 scale.
In German, M3.1 Flash (3.36) also leads every open model, but GPT-6 Luna is clearly ahead (3.62).
Sources: Artificial Analysis, MiniMax-M3.1-Flash and open-weights models (20 Oct 2026); Thorsten Meyer AI newsroom benchmark, 12 sources × English + German, run 20 Oct 2026.
Powered by Thorsten Meyer AI
Lower Prices Meet Performance Trade-offs
M3.1 Flash’s low listed token prices could appeal to developers running agent workflows, where a model may generate substantial amounts of text. But the reported gap between MiniMax’s speed claim and Artificial Analysis’s measurement, alongside the model’s above-median token use, shows why headline pricing alone may not predict the cost or speed of a completed task.
The model’s position also matters to teams comparing hosted AI services with models whose weights can be downloaded and run independently. For now, Artificial Analysis classifies M3.1 Flash as proprietary because the weights are unavailable. That limits what users can verify or deploy themselves, and makes the promised release and license terms material to any assessment of its openness.
AI language model performance comparison
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From Space Bunny to M3.1
Before MiniMax identified the model, it ran anonymously on OpenRouter as Space Bunny Alpha. ThorstenMeyerAI.com reports that it appeared on September 23, led OpenRouter’s usage ranking for 10 days, processed 56.7 trillion tokens there, and was removed on October 5. MiniMax says it processed more than 200 trillion tokens overall; the report does not provide further details about that total.
The anonymous period raises a data-handling question for users. According to the source, the provider could retain prompts during that period, although it was not allowed to train on them. MiniMax has not said what happened to the retained data. The new model has a one-million-token context window, adjustable reasoning effort, and an OpenRouter output limit listed at 524,288 tokens. OpenRouter offered it free for its first 72 hours.
The report also cautions against treating MiniMax’s benchmark chart as a full field comparison. It says the company chose the rivals, including GPT-5.6 Sol and Claude Opus 5, and omitted newer models named in the source. MiniMax’s chart gives M3.1 Flash 29.6 on Agents’ Last Exam, while DeepSeek’s model card reports 35.4 for DeepSeek V4.1 Flash 1017; the setups may differ, and DeepSeek is absent from that row in MiniMax’s chart.
““More aggressive sparsity.””
— MiniMax
Weights, Data and Test Results
The model’s weights had not been released when the source report was published, and MiniMax had not disclosed an official parameter count or license. One outlet’s reported parameter figures are explicitly unconfirmed and were not used by ThorstenMeyerAI.com. The source says weights were expected on October 23, but does not confirm that they were subsequently released.
The supplied report describes a test of 24 articles in English and German, based on 12 real news sources and judged by GPT-6.1 Sol and Claude Opus 5.5. It says each article was scored for source fidelity, invented facts, clarity, usefulness, structure and polish. However, the provided source material cuts off before the results table, so it does not establish how M3.1 Flash scored or how it compared with other models in that newsroom test.
It is also unclear what happened to prompts that could be retained during the Space Bunny Alpha period. MiniMax’s company benchmarks and internal security results have not been independently reproduced in the cited report. Those gaps limit conclusions about privacy, real-world performance and the company’s broader claims.
Watch for Weights and Scores
The next stated milestone is the promised October 23 weights release. Its timing, license and technical details would determine whether developers can inspect and run the model outside MiniMax’s services. Independent testing after release could also clarify whether performance changes across hosting setups.
Readers seeking the newsroom comparison need the missing results table or a complete version of the publisher’s test report. Until those figures are available, the evidence supports a narrower conclusion: M3.1 Flash is inexpensive by listed token price, but independent measurements show trade-offs in speed and token use, while its performance on the described journalism test remains unreported in the supplied material.
Key Questions
When did MiniMax release M3.1 Flash?
MiniMax released M3.1 Flash on October 19, 2026, according to ThorstenMeyerAI.com.
How much does M3.1 Flash cost?
The listed price is $0.15 per million input tokens and $0.60 per million output tokens. Cached input is listed at $0.015 per million tokens. MiniMax’s claim that it costs about 70% less per agent task has not been independently verified in the report.
How did it compare with other models?
Artificial Analysis gave it an Intelligence Index score of 43, below several listed open-weight models. MiniMax’s own benchmarks showed mixed results: M3.1 Flash led on Automation-Bench among the models shown, while DeepSeek V4.1 Flash 1017 scored higher on DeepSWE and Terminal-Bench.
Are M3.1 Flash’s weights available?
They were not available when the source report was published. MiniMax said the weights would arrive on October 23, but the supplied material does not confirm whether that release occurred or state the license.
What did the newsroom test find?
The report describes a test of 24 articles in two languages, scored by two AI judges, but the supplied source ends before the results table. It therefore does not provide the model’s newsroom-test scores or a basis for ranking it against the other tested models.
Source: ThorstenMeyerAI.com
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
